1 | /* |
2 | * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "compiler/compileLog.hpp" |
27 | #include "gc/shared/collectedHeap.inline.hpp" |
28 | #include "libadt/vectset.hpp" |
29 | #include "memory/universe.hpp" |
30 | #include "opto/addnode.hpp" |
31 | #include "opto/arraycopynode.hpp" |
32 | #include "opto/callnode.hpp" |
33 | #include "opto/castnode.hpp" |
34 | #include "opto/cfgnode.hpp" |
35 | #include "opto/compile.hpp" |
36 | #include "opto/convertnode.hpp" |
37 | #include "opto/graphKit.hpp" |
38 | #include "opto/locknode.hpp" |
39 | #include "opto/loopnode.hpp" |
40 | #include "opto/macro.hpp" |
41 | #include "opto/memnode.hpp" |
42 | #include "opto/narrowptrnode.hpp" |
43 | #include "opto/node.hpp" |
44 | #include "opto/opaquenode.hpp" |
45 | #include "opto/phaseX.hpp" |
46 | #include "opto/rootnode.hpp" |
47 | #include "opto/runtime.hpp" |
48 | #include "opto/subnode.hpp" |
49 | #include "opto/type.hpp" |
50 | #include "runtime/sharedRuntime.hpp" |
51 | #include "utilities/macros.hpp" |
52 | #if INCLUDE_G1GC |
53 | #include "gc/g1/g1ThreadLocalData.hpp" |
54 | #endif // INCLUDE_G1GC |
55 | #if INCLUDE_SHENANDOAHGC |
56 | #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" |
57 | #endif |
58 | |
59 | |
60 | // |
61 | // Replace any references to "oldref" in inputs to "use" with "newref". |
62 | // Returns the number of replacements made. |
63 | // |
64 | int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { |
65 | int nreplacements = 0; |
66 | uint req = use->req(); |
67 | for (uint j = 0; j < use->len(); j++) { |
68 | Node *uin = use->in(j); |
69 | if (uin == oldref) { |
70 | if (j < req) |
71 | use->set_req(j, newref); |
72 | else |
73 | use->set_prec(j, newref); |
74 | nreplacements++; |
75 | } else if (j >= req && uin == NULL) { |
76 | break; |
77 | } |
78 | } |
79 | return nreplacements; |
80 | } |
81 | |
82 | void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { |
83 | // Copy debug information and adjust JVMState information |
84 | uint old_dbg_start = oldcall->tf()->domain()->cnt(); |
85 | uint new_dbg_start = newcall->tf()->domain()->cnt(); |
86 | int jvms_adj = new_dbg_start - old_dbg_start; |
87 | assert (new_dbg_start == newcall->req(), "argument count mismatch" ); |
88 | |
89 | // SafePointScalarObject node could be referenced several times in debug info. |
90 | // Use Dict to record cloned nodes. |
91 | Dict* sosn_map = new Dict(cmpkey,hashkey); |
92 | for (uint i = old_dbg_start; i < oldcall->req(); i++) { |
93 | Node* old_in = oldcall->in(i); |
94 | // Clone old SafePointScalarObjectNodes, adjusting their field contents. |
95 | if (old_in != NULL && old_in->is_SafePointScalarObject()) { |
96 | SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); |
97 | uint old_unique = C->unique(); |
98 | Node* new_in = old_sosn->clone(sosn_map); |
99 | if (old_unique != C->unique()) { // New node? |
100 | new_in->set_req(0, C->root()); // reset control edge |
101 | new_in = transform_later(new_in); // Register new node. |
102 | } |
103 | old_in = new_in; |
104 | } |
105 | newcall->add_req(old_in); |
106 | } |
107 | |
108 | // JVMS may be shared so clone it before we modify it |
109 | newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL); |
110 | for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { |
111 | jvms->set_map(newcall); |
112 | jvms->set_locoff(jvms->locoff()+jvms_adj); |
113 | jvms->set_stkoff(jvms->stkoff()+jvms_adj); |
114 | jvms->set_monoff(jvms->monoff()+jvms_adj); |
115 | jvms->set_scloff(jvms->scloff()+jvms_adj); |
116 | jvms->set_endoff(jvms->endoff()+jvms_adj); |
117 | } |
118 | } |
119 | |
120 | Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { |
121 | Node* cmp; |
122 | if (mask != 0) { |
123 | Node* and_node = transform_later(new AndXNode(word, MakeConX(mask))); |
124 | cmp = transform_later(new CmpXNode(and_node, MakeConX(bits))); |
125 | } else { |
126 | cmp = word; |
127 | } |
128 | Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne)); |
129 | IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); |
130 | transform_later(iff); |
131 | |
132 | // Fast path taken. |
133 | Node *fast_taken = transform_later(new IfFalseNode(iff)); |
134 | |
135 | // Fast path not-taken, i.e. slow path |
136 | Node *slow_taken = transform_later(new IfTrueNode(iff)); |
137 | |
138 | if (return_fast_path) { |
139 | region->init_req(edge, slow_taken); // Capture slow-control |
140 | return fast_taken; |
141 | } else { |
142 | region->init_req(edge, fast_taken); // Capture fast-control |
143 | return slow_taken; |
144 | } |
145 | } |
146 | |
147 | //--------------------copy_predefined_input_for_runtime_call-------------------- |
148 | void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { |
149 | // Set fixed predefined input arguments |
150 | call->init_req( TypeFunc::Control, ctrl ); |
151 | call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); |
152 | call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? |
153 | call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); |
154 | call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); |
155 | } |
156 | |
157 | //------------------------------make_slow_call--------------------------------- |
158 | CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, |
159 | address slow_call, const char* leaf_name, Node* slow_path, |
160 | Node* parm0, Node* parm1, Node* parm2) { |
161 | |
162 | // Slow-path call |
163 | CallNode *call = leaf_name |
164 | ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) |
165 | : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); |
166 | |
167 | // Slow path call has no side-effects, uses few values |
168 | copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); |
169 | if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); |
170 | if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); |
171 | if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2); |
172 | copy_call_debug_info(oldcall, call); |
173 | call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. |
174 | _igvn.replace_node(oldcall, call); |
175 | transform_later(call); |
176 | |
177 | return call; |
178 | } |
179 | |
180 | void PhaseMacroExpand::extract_call_projections(CallNode *call) { |
181 | _fallthroughproj = NULL; |
182 | _fallthroughcatchproj = NULL; |
183 | _ioproj_fallthrough = NULL; |
184 | _ioproj_catchall = NULL; |
185 | _catchallcatchproj = NULL; |
186 | _memproj_fallthrough = NULL; |
187 | _memproj_catchall = NULL; |
188 | _resproj = NULL; |
189 | for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { |
190 | ProjNode *pn = call->fast_out(i)->as_Proj(); |
191 | switch (pn->_con) { |
192 | case TypeFunc::Control: |
193 | { |
194 | // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj |
195 | _fallthroughproj = pn; |
196 | DUIterator_Fast jmax, j = pn->fast_outs(jmax); |
197 | const Node *cn = pn->fast_out(j); |
198 | if (cn->is_Catch()) { |
199 | ProjNode *cpn = NULL; |
200 | for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { |
201 | cpn = cn->fast_out(k)->as_Proj(); |
202 | assert(cpn->is_CatchProj(), "must be a CatchProjNode" ); |
203 | if (cpn->_con == CatchProjNode::fall_through_index) |
204 | _fallthroughcatchproj = cpn; |
205 | else { |
206 | assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index." ); |
207 | _catchallcatchproj = cpn; |
208 | } |
209 | } |
210 | } |
211 | break; |
212 | } |
213 | case TypeFunc::I_O: |
214 | if (pn->_is_io_use) |
215 | _ioproj_catchall = pn; |
216 | else |
217 | _ioproj_fallthrough = pn; |
218 | break; |
219 | case TypeFunc::Memory: |
220 | if (pn->_is_io_use) |
221 | _memproj_catchall = pn; |
222 | else |
223 | _memproj_fallthrough = pn; |
224 | break; |
225 | case TypeFunc::Parms: |
226 | _resproj = pn; |
227 | break; |
228 | default: |
229 | assert(false, "unexpected projection from allocation node." ); |
230 | } |
231 | } |
232 | |
233 | } |
234 | |
235 | void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) { |
236 | BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2(); |
237 | bs->eliminate_gc_barrier(this, p2x); |
238 | } |
239 | |
240 | // Search for a memory operation for the specified memory slice. |
241 | static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { |
242 | Node *orig_mem = mem; |
243 | Node *alloc_mem = alloc->in(TypeFunc::Memory); |
244 | const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); |
245 | while (true) { |
246 | if (mem == alloc_mem || mem == start_mem ) { |
247 | return mem; // hit one of our sentinels |
248 | } else if (mem->is_MergeMem()) { |
249 | mem = mem->as_MergeMem()->memory_at(alias_idx); |
250 | } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { |
251 | Node *in = mem->in(0); |
252 | // we can safely skip over safepoints, calls, locks and membars because we |
253 | // already know that the object is safe to eliminate. |
254 | if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { |
255 | return in; |
256 | } else if (in->is_Call()) { |
257 | CallNode *call = in->as_Call(); |
258 | if (call->may_modify(tinst, phase)) { |
259 | assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape" ); |
260 | if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) { |
261 | return in; |
262 | } |
263 | } |
264 | mem = in->in(TypeFunc::Memory); |
265 | } else if (in->is_MemBar()) { |
266 | ArrayCopyNode* ac = NULL; |
267 | if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) { |
268 | assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone" ); |
269 | return ac; |
270 | } |
271 | mem = in->in(TypeFunc::Memory); |
272 | } else { |
273 | assert(false, "unexpected projection" ); |
274 | } |
275 | } else if (mem->is_Store()) { |
276 | const TypePtr* atype = mem->as_Store()->adr_type(); |
277 | int adr_idx = phase->C->get_alias_index(atype); |
278 | if (adr_idx == alias_idx) { |
279 | assert(atype->isa_oopptr(), "address type must be oopptr" ); |
280 | int adr_offset = atype->offset(); |
281 | uint adr_iid = atype->is_oopptr()->instance_id(); |
282 | // Array elements references have the same alias_idx |
283 | // but different offset and different instance_id. |
284 | if (adr_offset == offset && adr_iid == alloc->_idx) |
285 | return mem; |
286 | } else { |
287 | assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw" ); |
288 | } |
289 | mem = mem->in(MemNode::Memory); |
290 | } else if (mem->is_ClearArray()) { |
291 | if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { |
292 | // Can not bypass initialization of the instance |
293 | // we are looking. |
294 | debug_only(intptr_t offset;) |
295 | assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity" ); |
296 | InitializeNode* init = alloc->as_Allocate()->initialization(); |
297 | // We are looking for stored value, return Initialize node |
298 | // or memory edge from Allocate node. |
299 | if (init != NULL) |
300 | return init; |
301 | else |
302 | return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). |
303 | } |
304 | // Otherwise skip it (the call updated 'mem' value). |
305 | } else if (mem->Opcode() == Op_SCMemProj) { |
306 | mem = mem->in(0); |
307 | Node* adr = NULL; |
308 | if (mem->is_LoadStore()) { |
309 | adr = mem->in(MemNode::Address); |
310 | } else { |
311 | assert(mem->Opcode() == Op_EncodeISOArray || |
312 | mem->Opcode() == Op_StrCompressedCopy, "sanity" ); |
313 | adr = mem->in(3); // Destination array |
314 | } |
315 | const TypePtr* atype = adr->bottom_type()->is_ptr(); |
316 | int adr_idx = phase->C->get_alias_index(atype); |
317 | if (adr_idx == alias_idx) { |
318 | DEBUG_ONLY(mem->dump();) |
319 | assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field" ); |
320 | return NULL; |
321 | } |
322 | mem = mem->in(MemNode::Memory); |
323 | } else if (mem->Opcode() == Op_StrInflatedCopy) { |
324 | Node* adr = mem->in(3); // Destination array |
325 | const TypePtr* atype = adr->bottom_type()->is_ptr(); |
326 | int adr_idx = phase->C->get_alias_index(atype); |
327 | if (adr_idx == alias_idx) { |
328 | DEBUG_ONLY(mem->dump();) |
329 | assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field" ); |
330 | return NULL; |
331 | } |
332 | mem = mem->in(MemNode::Memory); |
333 | } else { |
334 | return mem; |
335 | } |
336 | assert(mem != orig_mem, "dead memory loop" ); |
337 | } |
338 | } |
339 | |
340 | // Generate loads from source of the arraycopy for fields of |
341 | // destination needed at a deoptimization point |
342 | Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) { |
343 | BasicType bt = ft; |
344 | const Type *type = ftype; |
345 | if (ft == T_NARROWOOP) { |
346 | bt = T_OBJECT; |
347 | type = ftype->make_oopptr(); |
348 | } |
349 | Node* res = NULL; |
350 | if (ac->is_clonebasic()) { |
351 | Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base); |
352 | Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset))); |
353 | const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); |
354 | res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); |
355 | } else { |
356 | if (ac->modifies(offset, offset, &_igvn, true)) { |
357 | assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result" ); |
358 | uint shift = exact_log2(type2aelembytes(bt)); |
359 | Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos))); |
360 | #ifdef _LP64 |
361 | diff = _igvn.transform(new ConvI2LNode(diff)); |
362 | #endif |
363 | diff = _igvn.transform(new LShiftXNode(diff, intcon(shift))); |
364 | |
365 | Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff)); |
366 | Node* base = ac->in(ArrayCopyNode::Src); |
367 | Node* adr = _igvn.transform(new AddPNode(base, base, off)); |
368 | const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset); |
369 | res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned); |
370 | } |
371 | } |
372 | if (res != NULL) { |
373 | res = _igvn.transform(res); |
374 | if (ftype->isa_narrowoop()) { |
375 | // PhaseMacroExpand::scalar_replacement adds DecodeN nodes |
376 | res = _igvn.transform(new EncodePNode(res, ftype)); |
377 | } |
378 | return res; |
379 | } |
380 | return NULL; |
381 | } |
382 | |
383 | // |
384 | // Given a Memory Phi, compute a value Phi containing the values from stores |
385 | // on the input paths. |
386 | // Note: this function is recursive, its depth is limited by the "level" argument |
387 | // Returns the computed Phi, or NULL if it cannot compute it. |
388 | Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) { |
389 | assert(mem->is_Phi(), "sanity" ); |
390 | int alias_idx = C->get_alias_index(adr_t); |
391 | int offset = adr_t->offset(); |
392 | int instance_id = adr_t->instance_id(); |
393 | |
394 | // Check if an appropriate value phi already exists. |
395 | Node* region = mem->in(0); |
396 | for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { |
397 | Node* phi = region->fast_out(k); |
398 | if (phi->is_Phi() && phi != mem && |
399 | phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) { |
400 | return phi; |
401 | } |
402 | } |
403 | // Check if an appropriate new value phi already exists. |
404 | Node* new_phi = value_phis->find(mem->_idx); |
405 | if (new_phi != NULL) |
406 | return new_phi; |
407 | |
408 | if (level <= 0) { |
409 | return NULL; // Give up: phi tree too deep |
410 | } |
411 | Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); |
412 | Node *alloc_mem = alloc->in(TypeFunc::Memory); |
413 | |
414 | uint length = mem->req(); |
415 | GrowableArray <Node *> values(length, length, NULL, false); |
416 | |
417 | // create a new Phi for the value |
418 | PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset); |
419 | transform_later(phi); |
420 | value_phis->push(phi, mem->_idx); |
421 | |
422 | for (uint j = 1; j < length; j++) { |
423 | Node *in = mem->in(j); |
424 | if (in == NULL || in->is_top()) { |
425 | values.at_put(j, in); |
426 | } else { |
427 | Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); |
428 | if (val == start_mem || val == alloc_mem) { |
429 | // hit a sentinel, return appropriate 0 value |
430 | values.at_put(j, _igvn.zerocon(ft)); |
431 | continue; |
432 | } |
433 | if (val->is_Initialize()) { |
434 | val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); |
435 | } |
436 | if (val == NULL) { |
437 | return NULL; // can't find a value on this path |
438 | } |
439 | if (val == mem) { |
440 | values.at_put(j, mem); |
441 | } else if (val->is_Store()) { |
442 | Node* n = val->in(MemNode::ValueIn); |
443 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
444 | n = bs->step_over_gc_barrier(n); |
445 | values.at_put(j, n); |
446 | } else if(val->is_Proj() && val->in(0) == alloc) { |
447 | values.at_put(j, _igvn.zerocon(ft)); |
448 | } else if (val->is_Phi()) { |
449 | val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); |
450 | if (val == NULL) { |
451 | return NULL; |
452 | } |
453 | values.at_put(j, val); |
454 | } else if (val->Opcode() == Op_SCMemProj) { |
455 | assert(val->in(0)->is_LoadStore() || |
456 | val->in(0)->Opcode() == Op_EncodeISOArray || |
457 | val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity" ); |
458 | assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field" ); |
459 | return NULL; |
460 | } else if (val->is_ArrayCopy()) { |
461 | Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc); |
462 | if (res == NULL) { |
463 | return NULL; |
464 | } |
465 | values.at_put(j, res); |
466 | } else { |
467 | #ifdef ASSERT |
468 | val->dump(); |
469 | assert(false, "unknown node on this path" ); |
470 | #endif |
471 | return NULL; // unknown node on this path |
472 | } |
473 | } |
474 | } |
475 | // Set Phi's inputs |
476 | for (uint j = 1; j < length; j++) { |
477 | if (values.at(j) == mem) { |
478 | phi->init_req(j, phi); |
479 | } else { |
480 | phi->init_req(j, values.at(j)); |
481 | } |
482 | } |
483 | return phi; |
484 | } |
485 | |
486 | // Search the last value stored into the object's field. |
487 | Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) { |
488 | assert(adr_t->is_known_instance_field(), "instance required" ); |
489 | int instance_id = adr_t->instance_id(); |
490 | assert((uint)instance_id == alloc->_idx, "wrong allocation" ); |
491 | |
492 | int alias_idx = C->get_alias_index(adr_t); |
493 | int offset = adr_t->offset(); |
494 | Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory); |
495 | Node *alloc_ctrl = alloc->in(TypeFunc::Control); |
496 | Node *alloc_mem = alloc->in(TypeFunc::Memory); |
497 | Arena *a = Thread::current()->resource_area(); |
498 | VectorSet visited(a); |
499 | |
500 | |
501 | bool done = sfpt_mem == alloc_mem; |
502 | Node *mem = sfpt_mem; |
503 | while (!done) { |
504 | if (visited.test_set(mem->_idx)) { |
505 | return NULL; // found a loop, give up |
506 | } |
507 | mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); |
508 | if (mem == start_mem || mem == alloc_mem) { |
509 | done = true; // hit a sentinel, return appropriate 0 value |
510 | } else if (mem->is_Initialize()) { |
511 | mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); |
512 | if (mem == NULL) { |
513 | done = true; // Something go wrong. |
514 | } else if (mem->is_Store()) { |
515 | const TypePtr* atype = mem->as_Store()->adr_type(); |
516 | assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice" ); |
517 | done = true; |
518 | } |
519 | } else if (mem->is_Store()) { |
520 | const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); |
521 | assert(atype != NULL, "address type must be oopptr" ); |
522 | assert(C->get_alias_index(atype) == alias_idx && |
523 | atype->is_known_instance_field() && atype->offset() == offset && |
524 | atype->instance_id() == instance_id, "store is correct memory slice" ); |
525 | done = true; |
526 | } else if (mem->is_Phi()) { |
527 | // try to find a phi's unique input |
528 | Node *unique_input = NULL; |
529 | Node *top = C->top(); |
530 | for (uint i = 1; i < mem->req(); i++) { |
531 | Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); |
532 | if (n == NULL || n == top || n == mem) { |
533 | continue; |
534 | } else if (unique_input == NULL) { |
535 | unique_input = n; |
536 | } else if (unique_input != n) { |
537 | unique_input = top; |
538 | break; |
539 | } |
540 | } |
541 | if (unique_input != NULL && unique_input != top) { |
542 | mem = unique_input; |
543 | } else { |
544 | done = true; |
545 | } |
546 | } else if (mem->is_ArrayCopy()) { |
547 | done = true; |
548 | } else { |
549 | assert(false, "unexpected node" ); |
550 | } |
551 | } |
552 | if (mem != NULL) { |
553 | if (mem == start_mem || mem == alloc_mem) { |
554 | // hit a sentinel, return appropriate 0 value |
555 | return _igvn.zerocon(ft); |
556 | } else if (mem->is_Store()) { |
557 | Node* n = mem->in(MemNode::ValueIn); |
558 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
559 | n = bs->step_over_gc_barrier(n); |
560 | return n; |
561 | } else if (mem->is_Phi()) { |
562 | // attempt to produce a Phi reflecting the values on the input paths of the Phi |
563 | Node_Stack value_phis(a, 8); |
564 | Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); |
565 | if (phi != NULL) { |
566 | return phi; |
567 | } else { |
568 | // Kill all new Phis |
569 | while(value_phis.is_nonempty()) { |
570 | Node* n = value_phis.node(); |
571 | _igvn.replace_node(n, C->top()); |
572 | value_phis.pop(); |
573 | } |
574 | } |
575 | } else if (mem->is_ArrayCopy()) { |
576 | Node* ctl = mem->in(0); |
577 | Node* m = mem->in(TypeFunc::Memory); |
578 | if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) { |
579 | // pin the loads in the uncommon trap path |
580 | ctl = sfpt_ctl; |
581 | m = sfpt_mem; |
582 | } |
583 | return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc); |
584 | } |
585 | } |
586 | // Something go wrong. |
587 | return NULL; |
588 | } |
589 | |
590 | // Check the possibility of scalar replacement. |
591 | bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { |
592 | // Scan the uses of the allocation to check for anything that would |
593 | // prevent us from eliminating it. |
594 | NOT_PRODUCT( const char* fail_eliminate = NULL; ) |
595 | DEBUG_ONLY( Node* disq_node = NULL; ) |
596 | bool can_eliminate = true; |
597 | |
598 | Node* res = alloc->result_cast(); |
599 | const TypeOopPtr* res_type = NULL; |
600 | if (res == NULL) { |
601 | // All users were eliminated. |
602 | } else if (!res->is_CheckCastPP()) { |
603 | NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP" ;) |
604 | can_eliminate = false; |
605 | } else { |
606 | res_type = _igvn.type(res)->isa_oopptr(); |
607 | if (res_type == NULL) { |
608 | NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation" ;) |
609 | can_eliminate = false; |
610 | } else if (res_type->isa_aryptr()) { |
611 | int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); |
612 | if (length < 0) { |
613 | NOT_PRODUCT(fail_eliminate = "Array's size is not constant" ;) |
614 | can_eliminate = false; |
615 | } |
616 | } |
617 | } |
618 | |
619 | if (can_eliminate && res != NULL) { |
620 | for (DUIterator_Fast jmax, j = res->fast_outs(jmax); |
621 | j < jmax && can_eliminate; j++) { |
622 | Node* use = res->fast_out(j); |
623 | |
624 | if (use->is_AddP()) { |
625 | const TypePtr* addp_type = _igvn.type(use)->is_ptr(); |
626 | int offset = addp_type->offset(); |
627 | |
628 | if (offset == Type::OffsetTop || offset == Type::OffsetBot) { |
629 | NOT_PRODUCT(fail_eliminate = "Undefined field referrence" ;) |
630 | can_eliminate = false; |
631 | break; |
632 | } |
633 | for (DUIterator_Fast kmax, k = use->fast_outs(kmax); |
634 | k < kmax && can_eliminate; k++) { |
635 | Node* n = use->fast_out(k); |
636 | if (!n->is_Store() && n->Opcode() != Op_CastP2X && |
637 | SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&) |
638 | !(n->is_ArrayCopy() && |
639 | n->as_ArrayCopy()->is_clonebasic() && |
640 | n->in(ArrayCopyNode::Dest) == use)) { |
641 | DEBUG_ONLY(disq_node = n;) |
642 | if (n->is_Load() || n->is_LoadStore()) { |
643 | NOT_PRODUCT(fail_eliminate = "Field load" ;) |
644 | } else { |
645 | NOT_PRODUCT(fail_eliminate = "Not store field referrence" ;) |
646 | } |
647 | can_eliminate = false; |
648 | } |
649 | } |
650 | } else if (use->is_ArrayCopy() && |
651 | (use->as_ArrayCopy()->is_arraycopy_validated() || |
652 | use->as_ArrayCopy()->is_copyof_validated() || |
653 | use->as_ArrayCopy()->is_copyofrange_validated()) && |
654 | use->in(ArrayCopyNode::Dest) == res) { |
655 | // ok to eliminate |
656 | } else if (use->is_SafePoint()) { |
657 | SafePointNode* sfpt = use->as_SafePoint(); |
658 | if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { |
659 | // Object is passed as argument. |
660 | DEBUG_ONLY(disq_node = use;) |
661 | NOT_PRODUCT(fail_eliminate = "Object is passed as argument" ;) |
662 | can_eliminate = false; |
663 | } |
664 | Node* sfptMem = sfpt->memory(); |
665 | if (sfptMem == NULL || sfptMem->is_top()) { |
666 | DEBUG_ONLY(disq_node = use;) |
667 | NOT_PRODUCT(fail_eliminate = "NULL or TOP memory" ;) |
668 | can_eliminate = false; |
669 | } else { |
670 | safepoints.append_if_missing(sfpt); |
671 | } |
672 | } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark |
673 | if (use->is_Phi()) { |
674 | if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { |
675 | NOT_PRODUCT(fail_eliminate = "Object is return value" ;) |
676 | } else { |
677 | NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi" ;) |
678 | } |
679 | DEBUG_ONLY(disq_node = use;) |
680 | } else { |
681 | if (use->Opcode() == Op_Return) { |
682 | NOT_PRODUCT(fail_eliminate = "Object is return value" ;) |
683 | }else { |
684 | NOT_PRODUCT(fail_eliminate = "Object is referenced by node" ;) |
685 | } |
686 | DEBUG_ONLY(disq_node = use;) |
687 | } |
688 | can_eliminate = false; |
689 | } |
690 | } |
691 | } |
692 | |
693 | #ifndef PRODUCT |
694 | if (PrintEliminateAllocations) { |
695 | if (can_eliminate) { |
696 | tty->print("Scalar " ); |
697 | if (res == NULL) |
698 | alloc->dump(); |
699 | else |
700 | res->dump(); |
701 | } else if (alloc->_is_scalar_replaceable) { |
702 | tty->print("NotScalar (%s)" , fail_eliminate); |
703 | if (res == NULL) |
704 | alloc->dump(); |
705 | else |
706 | res->dump(); |
707 | #ifdef ASSERT |
708 | if (disq_node != NULL) { |
709 | tty->print(" >>>> " ); |
710 | disq_node->dump(); |
711 | } |
712 | #endif /*ASSERT*/ |
713 | } |
714 | } |
715 | #endif |
716 | return can_eliminate; |
717 | } |
718 | |
719 | // Do scalar replacement. |
720 | bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { |
721 | GrowableArray <SafePointNode *> safepoints_done; |
722 | |
723 | ciKlass* klass = NULL; |
724 | ciInstanceKlass* iklass = NULL; |
725 | int nfields = 0; |
726 | int array_base = 0; |
727 | int element_size = 0; |
728 | BasicType basic_elem_type = T_ILLEGAL; |
729 | ciType* elem_type = NULL; |
730 | |
731 | Node* res = alloc->result_cast(); |
732 | assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result" ); |
733 | const TypeOopPtr* res_type = NULL; |
734 | if (res != NULL) { // Could be NULL when there are no users |
735 | res_type = _igvn.type(res)->isa_oopptr(); |
736 | } |
737 | |
738 | if (res != NULL) { |
739 | klass = res_type->klass(); |
740 | if (res_type->isa_instptr()) { |
741 | // find the fields of the class which will be needed for safepoint debug information |
742 | assert(klass->is_instance_klass(), "must be an instance klass." ); |
743 | iklass = klass->as_instance_klass(); |
744 | nfields = iklass->nof_nonstatic_fields(); |
745 | } else { |
746 | // find the array's elements which will be needed for safepoint debug information |
747 | nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); |
748 | assert(klass->is_array_klass() && nfields >= 0, "must be an array klass." ); |
749 | elem_type = klass->as_array_klass()->element_type(); |
750 | basic_elem_type = elem_type->basic_type(); |
751 | array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); |
752 | element_size = type2aelembytes(basic_elem_type); |
753 | } |
754 | } |
755 | // |
756 | // Process the safepoint uses |
757 | // |
758 | while (safepoints.length() > 0) { |
759 | SafePointNode* sfpt = safepoints.pop(); |
760 | Node* mem = sfpt->memory(); |
761 | Node* ctl = sfpt->control(); |
762 | assert(sfpt->jvms() != NULL, "missed JVMS" ); |
763 | // Fields of scalar objs are referenced only at the end |
764 | // of regular debuginfo at the last (youngest) JVMS. |
765 | // Record relative start index. |
766 | uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); |
767 | SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, |
768 | #ifdef ASSERT |
769 | alloc, |
770 | #endif |
771 | first_ind, nfields); |
772 | sobj->init_req(0, C->root()); |
773 | transform_later(sobj); |
774 | |
775 | // Scan object's fields adding an input to the safepoint for each field. |
776 | for (int j = 0; j < nfields; j++) { |
777 | intptr_t offset; |
778 | ciField* field = NULL; |
779 | if (iklass != NULL) { |
780 | field = iklass->nonstatic_field_at(j); |
781 | offset = field->offset(); |
782 | elem_type = field->type(); |
783 | basic_elem_type = field->layout_type(); |
784 | } else { |
785 | offset = array_base + j * (intptr_t)element_size; |
786 | } |
787 | |
788 | const Type *field_type; |
789 | // The next code is taken from Parse::do_get_xxx(). |
790 | if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { |
791 | if (!elem_type->is_loaded()) { |
792 | field_type = TypeInstPtr::BOTTOM; |
793 | } else if (field != NULL && field->is_static_constant()) { |
794 | // This can happen if the constant oop is non-perm. |
795 | ciObject* con = field->constant_value().as_object(); |
796 | // Do not "join" in the previous type; it doesn't add value, |
797 | // and may yield a vacuous result if the field is of interface type. |
798 | field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); |
799 | assert(field_type != NULL, "field singleton type must be consistent" ); |
800 | } else { |
801 | field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); |
802 | } |
803 | if (UseCompressedOops) { |
804 | field_type = field_type->make_narrowoop(); |
805 | basic_elem_type = T_NARROWOOP; |
806 | } |
807 | } else { |
808 | field_type = Type::get_const_basic_type(basic_elem_type); |
809 | } |
810 | |
811 | const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); |
812 | |
813 | Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc); |
814 | if (field_val == NULL) { |
815 | // We weren't able to find a value for this field, |
816 | // give up on eliminating this allocation. |
817 | |
818 | // Remove any extra entries we added to the safepoint. |
819 | uint last = sfpt->req() - 1; |
820 | for (int k = 0; k < j; k++) { |
821 | sfpt->del_req(last--); |
822 | } |
823 | _igvn._worklist.push(sfpt); |
824 | // rollback processed safepoints |
825 | while (safepoints_done.length() > 0) { |
826 | SafePointNode* sfpt_done = safepoints_done.pop(); |
827 | // remove any extra entries we added to the safepoint |
828 | last = sfpt_done->req() - 1; |
829 | for (int k = 0; k < nfields; k++) { |
830 | sfpt_done->del_req(last--); |
831 | } |
832 | JVMState *jvms = sfpt_done->jvms(); |
833 | jvms->set_endoff(sfpt_done->req()); |
834 | // Now make a pass over the debug information replacing any references |
835 | // to SafePointScalarObjectNode with the allocated object. |
836 | int start = jvms->debug_start(); |
837 | int end = jvms->debug_end(); |
838 | for (int i = start; i < end; i++) { |
839 | if (sfpt_done->in(i)->is_SafePointScalarObject()) { |
840 | SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); |
841 | if (scobj->first_index(jvms) == sfpt_done->req() && |
842 | scobj->n_fields() == (uint)nfields) { |
843 | assert(scobj->alloc() == alloc, "sanity" ); |
844 | sfpt_done->set_req(i, res); |
845 | } |
846 | } |
847 | } |
848 | _igvn._worklist.push(sfpt_done); |
849 | } |
850 | #ifndef PRODUCT |
851 | if (PrintEliminateAllocations) { |
852 | if (field != NULL) { |
853 | tty->print("=== At SafePoint node %d can't find value of Field: " , |
854 | sfpt->_idx); |
855 | field->print(); |
856 | int field_idx = C->get_alias_index(field_addr_type); |
857 | tty->print(" (alias_idx=%d)" , field_idx); |
858 | } else { // Array's element |
859 | tty->print("=== At SafePoint node %d can't find value of array element [%d]" , |
860 | sfpt->_idx, j); |
861 | } |
862 | tty->print(", which prevents elimination of: " ); |
863 | if (res == NULL) |
864 | alloc->dump(); |
865 | else |
866 | res->dump(); |
867 | } |
868 | #endif |
869 | return false; |
870 | } |
871 | if (UseCompressedOops && field_type->isa_narrowoop()) { |
872 | // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation |
873 | // to be able scalar replace the allocation. |
874 | if (field_val->is_EncodeP()) { |
875 | field_val = field_val->in(1); |
876 | } else { |
877 | field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type())); |
878 | } |
879 | } |
880 | sfpt->add_req(field_val); |
881 | } |
882 | JVMState *jvms = sfpt->jvms(); |
883 | jvms->set_endoff(sfpt->req()); |
884 | // Now make a pass over the debug information replacing any references |
885 | // to the allocated object with "sobj" |
886 | int start = jvms->debug_start(); |
887 | int end = jvms->debug_end(); |
888 | sfpt->replace_edges_in_range(res, sobj, start, end); |
889 | _igvn._worklist.push(sfpt); |
890 | safepoints_done.append_if_missing(sfpt); // keep it for rollback |
891 | } |
892 | return true; |
893 | } |
894 | |
895 | static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) { |
896 | Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control); |
897 | Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory); |
898 | if (ctl_proj != NULL) { |
899 | igvn.replace_node(ctl_proj, n->in(0)); |
900 | } |
901 | if (mem_proj != NULL) { |
902 | igvn.replace_node(mem_proj, n->in(TypeFunc::Memory)); |
903 | } |
904 | } |
905 | |
906 | // Process users of eliminated allocation. |
907 | void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { |
908 | Node* res = alloc->result_cast(); |
909 | if (res != NULL) { |
910 | for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { |
911 | Node *use = res->last_out(j); |
912 | uint oc1 = res->outcnt(); |
913 | |
914 | if (use->is_AddP()) { |
915 | for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { |
916 | Node *n = use->last_out(k); |
917 | uint oc2 = use->outcnt(); |
918 | if (n->is_Store()) { |
919 | #ifdef ASSERT |
920 | // Verify that there is no dependent MemBarVolatile nodes, |
921 | // they should be removed during IGVN, see MemBarNode::Ideal(). |
922 | for (DUIterator_Fast pmax, p = n->fast_outs(pmax); |
923 | p < pmax; p++) { |
924 | Node* mb = n->fast_out(p); |
925 | assert(mb->is_Initialize() || !mb->is_MemBar() || |
926 | mb->req() <= MemBarNode::Precedent || |
927 | mb->in(MemBarNode::Precedent) != n, |
928 | "MemBarVolatile should be eliminated for non-escaping object" ); |
929 | } |
930 | #endif |
931 | _igvn.replace_node(n, n->in(MemNode::Memory)); |
932 | } else if (n->is_ArrayCopy()) { |
933 | // Disconnect ArrayCopy node |
934 | ArrayCopyNode* ac = n->as_ArrayCopy(); |
935 | assert(ac->is_clonebasic(), "unexpected array copy kind" ); |
936 | Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out(); |
937 | disconnect_projections(ac, _igvn); |
938 | assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation" ); |
939 | Node* membar_before = alloc->in(0)->in(0); |
940 | disconnect_projections(membar_before->as_MemBar(), _igvn); |
941 | if (membar_after->is_MemBar()) { |
942 | disconnect_projections(membar_after->as_MemBar(), _igvn); |
943 | } |
944 | } else { |
945 | eliminate_gc_barrier(n); |
946 | } |
947 | k -= (oc2 - use->outcnt()); |
948 | } |
949 | _igvn.remove_dead_node(use); |
950 | } else if (use->is_ArrayCopy()) { |
951 | // Disconnect ArrayCopy node |
952 | ArrayCopyNode* ac = use->as_ArrayCopy(); |
953 | assert(ac->is_arraycopy_validated() || |
954 | ac->is_copyof_validated() || |
955 | ac->is_copyofrange_validated(), "unsupported" ); |
956 | CallProjections callprojs; |
957 | ac->extract_projections(&callprojs, true); |
958 | |
959 | _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O)); |
960 | _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory)); |
961 | _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control)); |
962 | |
963 | // Set control to top. IGVN will remove the remaining projections |
964 | ac->set_req(0, top()); |
965 | ac->replace_edge(res, top()); |
966 | |
967 | // Disconnect src right away: it can help find new |
968 | // opportunities for allocation elimination |
969 | Node* src = ac->in(ArrayCopyNode::Src); |
970 | ac->replace_edge(src, top()); |
971 | // src can be top at this point if src and dest of the |
972 | // arraycopy were the same |
973 | if (src->outcnt() == 0 && !src->is_top()) { |
974 | _igvn.remove_dead_node(src); |
975 | } |
976 | |
977 | _igvn._worklist.push(ac); |
978 | } else { |
979 | eliminate_gc_barrier(use); |
980 | } |
981 | j -= (oc1 - res->outcnt()); |
982 | } |
983 | assert(res->outcnt() == 0, "all uses of allocated objects must be deleted" ); |
984 | _igvn.remove_dead_node(res); |
985 | } |
986 | |
987 | // |
988 | // Process other users of allocation's projections |
989 | // |
990 | if (_resproj != NULL && _resproj->outcnt() != 0) { |
991 | // First disconnect stores captured by Initialize node. |
992 | // If Initialize node is eliminated first in the following code, |
993 | // it will kill such stores and DUIterator_Last will assert. |
994 | for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { |
995 | Node *use = _resproj->fast_out(j); |
996 | if (use->is_AddP()) { |
997 | // raw memory addresses used only by the initialization |
998 | _igvn.replace_node(use, C->top()); |
999 | --j; --jmax; |
1000 | } |
1001 | } |
1002 | for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { |
1003 | Node *use = _resproj->last_out(j); |
1004 | uint oc1 = _resproj->outcnt(); |
1005 | if (use->is_Initialize()) { |
1006 | // Eliminate Initialize node. |
1007 | InitializeNode *init = use->as_Initialize(); |
1008 | assert(init->outcnt() <= 2, "only a control and memory projection expected" ); |
1009 | Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control); |
1010 | if (ctrl_proj != NULL) { |
1011 | _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control)); |
1012 | #ifdef ASSERT |
1013 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
1014 | Node* tmp = init->in(TypeFunc::Control); |
1015 | while (bs->is_gc_barrier_node(tmp)) { |
1016 | Node* tmp2 = bs->step_over_gc_barrier_ctrl(tmp); |
1017 | assert(tmp != tmp2, "Must make progress" ); |
1018 | tmp = tmp2; |
1019 | } |
1020 | assert(tmp == _fallthroughcatchproj, "allocation control projection" ); |
1021 | #endif |
1022 | } |
1023 | Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory); |
1024 | if (mem_proj != NULL) { |
1025 | Node *mem = init->in(TypeFunc::Memory); |
1026 | #ifdef ASSERT |
1027 | if (mem->is_MergeMem()) { |
1028 | assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection" ); |
1029 | } else { |
1030 | assert(mem == _memproj_fallthrough, "allocation memory projection" ); |
1031 | } |
1032 | #endif |
1033 | _igvn.replace_node(mem_proj, mem); |
1034 | } |
1035 | } else { |
1036 | assert(false, "only Initialize or AddP expected" ); |
1037 | } |
1038 | j -= (oc1 - _resproj->outcnt()); |
1039 | } |
1040 | } |
1041 | if (_fallthroughcatchproj != NULL) { |
1042 | _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); |
1043 | } |
1044 | if (_memproj_fallthrough != NULL) { |
1045 | _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); |
1046 | } |
1047 | if (_memproj_catchall != NULL) { |
1048 | _igvn.replace_node(_memproj_catchall, C->top()); |
1049 | } |
1050 | if (_ioproj_fallthrough != NULL) { |
1051 | _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); |
1052 | } |
1053 | if (_ioproj_catchall != NULL) { |
1054 | _igvn.replace_node(_ioproj_catchall, C->top()); |
1055 | } |
1056 | if (_catchallcatchproj != NULL) { |
1057 | _igvn.replace_node(_catchallcatchproj, C->top()); |
1058 | } |
1059 | } |
1060 | |
1061 | bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { |
1062 | // Don't do scalar replacement if the frame can be popped by JVMTI: |
1063 | // if reallocation fails during deoptimization we'll pop all |
1064 | // interpreter frames for this compiled frame and that won't play |
1065 | // nice with JVMTI popframe. |
1066 | if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) { |
1067 | return false; |
1068 | } |
1069 | Node* klass = alloc->in(AllocateNode::KlassNode); |
1070 | const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); |
1071 | Node* res = alloc->result_cast(); |
1072 | // Eliminate boxing allocations which are not used |
1073 | // regardless scalar replacable status. |
1074 | bool boxing_alloc = C->eliminate_boxing() && |
1075 | tklass->klass()->is_instance_klass() && |
1076 | tklass->klass()->as_instance_klass()->is_box_klass(); |
1077 | if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { |
1078 | return false; |
1079 | } |
1080 | |
1081 | extract_call_projections(alloc); |
1082 | |
1083 | GrowableArray <SafePointNode *> safepoints; |
1084 | if (!can_eliminate_allocation(alloc, safepoints)) { |
1085 | return false; |
1086 | } |
1087 | |
1088 | if (!alloc->_is_scalar_replaceable) { |
1089 | assert(res == NULL, "sanity" ); |
1090 | // We can only eliminate allocation if all debug info references |
1091 | // are already replaced with SafePointScalarObject because |
1092 | // we can't search for a fields value without instance_id. |
1093 | if (safepoints.length() > 0) { |
1094 | return false; |
1095 | } |
1096 | } |
1097 | |
1098 | if (!scalar_replacement(alloc, safepoints)) { |
1099 | return false; |
1100 | } |
1101 | |
1102 | CompileLog* log = C->log(); |
1103 | if (log != NULL) { |
1104 | log->head("eliminate_allocation type='%d'" , |
1105 | log->identify(tklass->klass())); |
1106 | JVMState* p = alloc->jvms(); |
1107 | while (p != NULL) { |
1108 | log->elem("jvms bci='%d' method='%d'" , p->bci(), log->identify(p->method())); |
1109 | p = p->caller(); |
1110 | } |
1111 | log->tail("eliminate_allocation" ); |
1112 | } |
1113 | |
1114 | process_users_of_allocation(alloc); |
1115 | |
1116 | #ifndef PRODUCT |
1117 | if (PrintEliminateAllocations) { |
1118 | if (alloc->is_AllocateArray()) |
1119 | tty->print_cr("++++ Eliminated: %d AllocateArray" , alloc->_idx); |
1120 | else |
1121 | tty->print_cr("++++ Eliminated: %d Allocate" , alloc->_idx); |
1122 | } |
1123 | #endif |
1124 | |
1125 | return true; |
1126 | } |
1127 | |
1128 | bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { |
1129 | // EA should remove all uses of non-escaping boxing node. |
1130 | if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) { |
1131 | return false; |
1132 | } |
1133 | |
1134 | assert(boxing->result_cast() == NULL, "unexpected boxing node result" ); |
1135 | |
1136 | extract_call_projections(boxing); |
1137 | |
1138 | const TypeTuple* r = boxing->tf()->range(); |
1139 | assert(r->cnt() > TypeFunc::Parms, "sanity" ); |
1140 | const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); |
1141 | assert(t != NULL, "sanity" ); |
1142 | |
1143 | CompileLog* log = C->log(); |
1144 | if (log != NULL) { |
1145 | log->head("eliminate_boxing type='%d'" , |
1146 | log->identify(t->klass())); |
1147 | JVMState* p = boxing->jvms(); |
1148 | while (p != NULL) { |
1149 | log->elem("jvms bci='%d' method='%d'" , p->bci(), log->identify(p->method())); |
1150 | p = p->caller(); |
1151 | } |
1152 | log->tail("eliminate_boxing" ); |
1153 | } |
1154 | |
1155 | process_users_of_allocation(boxing); |
1156 | |
1157 | #ifndef PRODUCT |
1158 | if (PrintEliminateAllocations) { |
1159 | tty->print("++++ Eliminated: %d " , boxing->_idx); |
1160 | boxing->method()->print_short_name(tty); |
1161 | tty->cr(); |
1162 | } |
1163 | #endif |
1164 | |
1165 | return true; |
1166 | } |
1167 | |
1168 | //---------------------------set_eden_pointers------------------------- |
1169 | void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { |
1170 | if (UseTLAB) { // Private allocation: load from TLS |
1171 | Node* thread = transform_later(new ThreadLocalNode()); |
1172 | int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); |
1173 | int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); |
1174 | eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); |
1175 | eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); |
1176 | } else { // Shared allocation: load from globals |
1177 | CollectedHeap* ch = Universe::heap(); |
1178 | address top_adr = (address)ch->top_addr(); |
1179 | address end_adr = (address)ch->end_addr(); |
1180 | eden_top_adr = makecon(TypeRawPtr::make(top_adr)); |
1181 | eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); |
1182 | } |
1183 | } |
1184 | |
1185 | |
1186 | Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { |
1187 | Node* adr = basic_plus_adr(base, offset); |
1188 | const TypePtr* adr_type = adr->bottom_type()->is_ptr(); |
1189 | Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); |
1190 | transform_later(value); |
1191 | return value; |
1192 | } |
1193 | |
1194 | |
1195 | Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { |
1196 | Node* adr = basic_plus_adr(base, offset); |
1197 | mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); |
1198 | transform_later(mem); |
1199 | return mem; |
1200 | } |
1201 | |
1202 | //============================================================================= |
1203 | // |
1204 | // A L L O C A T I O N |
1205 | // |
1206 | // Allocation attempts to be fast in the case of frequent small objects. |
1207 | // It breaks down like this: |
1208 | // |
1209 | // 1) Size in doublewords is computed. This is a constant for objects and |
1210 | // variable for most arrays. Doubleword units are used to avoid size |
1211 | // overflow of huge doubleword arrays. We need doublewords in the end for |
1212 | // rounding. |
1213 | // |
1214 | // 2) Size is checked for being 'too large'. Too-large allocations will go |
1215 | // the slow path into the VM. The slow path can throw any required |
1216 | // exceptions, and does all the special checks for very large arrays. The |
1217 | // size test can constant-fold away for objects. For objects with |
1218 | // finalizers it constant-folds the otherway: you always go slow with |
1219 | // finalizers. |
1220 | // |
1221 | // 3) If NOT using TLABs, this is the contended loop-back point. |
1222 | // Load-Locked the heap top. If using TLABs normal-load the heap top. |
1223 | // |
1224 | // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. |
1225 | // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish |
1226 | // "size*8" we always enter the VM, where "largish" is a constant picked small |
1227 | // enough that there's always space between the eden max and 4Gig (old space is |
1228 | // there so it's quite large) and large enough that the cost of entering the VM |
1229 | // is dwarfed by the cost to initialize the space. |
1230 | // |
1231 | // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back |
1232 | // down. If contended, repeat at step 3. If using TLABs normal-store |
1233 | // adjusted heap top back down; there is no contention. |
1234 | // |
1235 | // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark |
1236 | // fields. |
1237 | // |
1238 | // 7) Merge with the slow-path; cast the raw memory pointer to the correct |
1239 | // oop flavor. |
1240 | // |
1241 | //============================================================================= |
1242 | // FastAllocateSizeLimit value is in DOUBLEWORDS. |
1243 | // Allocations bigger than this always go the slow route. |
1244 | // This value must be small enough that allocation attempts that need to |
1245 | // trigger exceptions go the slow route. Also, it must be small enough so |
1246 | // that heap_top + size_in_bytes does not wrap around the 4Gig limit. |
1247 | //=============================================================================j// |
1248 | // %%% Here is an old comment from parseHelper.cpp; is it outdated? |
1249 | // The allocator will coalesce int->oop copies away. See comment in |
1250 | // coalesce.cpp about how this works. It depends critically on the exact |
1251 | // code shape produced here, so if you are changing this code shape |
1252 | // make sure the GC info for the heap-top is correct in and around the |
1253 | // slow-path call. |
1254 | // |
1255 | |
1256 | void PhaseMacroExpand::expand_allocate_common( |
1257 | AllocateNode* alloc, // allocation node to be expanded |
1258 | Node* length, // array length for an array allocation |
1259 | const TypeFunc* slow_call_type, // Type of slow call |
1260 | address slow_call_address // Address of slow call |
1261 | ) |
1262 | { |
1263 | |
1264 | Node* ctrl = alloc->in(TypeFunc::Control); |
1265 | Node* mem = alloc->in(TypeFunc::Memory); |
1266 | Node* i_o = alloc->in(TypeFunc::I_O); |
1267 | Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); |
1268 | Node* klass_node = alloc->in(AllocateNode::KlassNode); |
1269 | Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); |
1270 | |
1271 | assert(ctrl != NULL, "must have control" ); |
1272 | // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. |
1273 | // they will not be used if "always_slow" is set |
1274 | enum { slow_result_path = 1, fast_result_path = 2 }; |
1275 | Node *result_region = NULL; |
1276 | Node *result_phi_rawmem = NULL; |
1277 | Node *result_phi_rawoop = NULL; |
1278 | Node *result_phi_i_o = NULL; |
1279 | |
1280 | // The initial slow comparison is a size check, the comparison |
1281 | // we want to do is a BoolTest::gt |
1282 | bool always_slow = false; |
1283 | int tv = _igvn.find_int_con(initial_slow_test, -1); |
1284 | if (tv >= 0) { |
1285 | always_slow = (tv == 1); |
1286 | initial_slow_test = NULL; |
1287 | } else { |
1288 | initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); |
1289 | } |
1290 | |
1291 | if (C->env()->dtrace_alloc_probes() || |
1292 | (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) { |
1293 | // Force slow-path allocation |
1294 | always_slow = true; |
1295 | initial_slow_test = NULL; |
1296 | } |
1297 | |
1298 | |
1299 | enum { too_big_or_final_path = 1, need_gc_path = 2 }; |
1300 | Node *slow_region = NULL; |
1301 | Node *toobig_false = ctrl; |
1302 | |
1303 | assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent" ); |
1304 | // generate the initial test if necessary |
1305 | if (initial_slow_test != NULL ) { |
1306 | slow_region = new RegionNode(3); |
1307 | |
1308 | // Now make the initial failure test. Usually a too-big test but |
1309 | // might be a TRUE for finalizers or a fancy class check for |
1310 | // newInstance0. |
1311 | IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); |
1312 | transform_later(toobig_iff); |
1313 | // Plug the failing-too-big test into the slow-path region |
1314 | Node *toobig_true = new IfTrueNode( toobig_iff ); |
1315 | transform_later(toobig_true); |
1316 | slow_region ->init_req( too_big_or_final_path, toobig_true ); |
1317 | toobig_false = new IfFalseNode( toobig_iff ); |
1318 | transform_later(toobig_false); |
1319 | } else { // No initial test, just fall into next case |
1320 | toobig_false = ctrl; |
1321 | debug_only(slow_region = NodeSentinel); |
1322 | } |
1323 | |
1324 | Node *slow_mem = mem; // save the current memory state for slow path |
1325 | // generate the fast allocation code unless we know that the initial test will always go slow |
1326 | if (!always_slow) { |
1327 | // Fast path modifies only raw memory. |
1328 | if (mem->is_MergeMem()) { |
1329 | mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); |
1330 | } |
1331 | |
1332 | // allocate the Region and Phi nodes for the result |
1333 | result_region = new RegionNode(3); |
1334 | result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); |
1335 | result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM); |
1336 | result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch |
1337 | |
1338 | // Grab regular I/O before optional prefetch may change it. |
1339 | // Slow-path does no I/O so just set it to the original I/O. |
1340 | result_phi_i_o->init_req(slow_result_path, i_o); |
1341 | |
1342 | Node* needgc_ctrl = NULL; |
1343 | // Name successful fast-path variables |
1344 | Node* fast_oop_ctrl; |
1345 | Node* fast_oop_rawmem; |
1346 | |
1347 | intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines; |
1348 | |
1349 | BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); |
1350 | Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl, |
1351 | fast_oop_ctrl, fast_oop_rawmem, |
1352 | prefetch_lines); |
1353 | |
1354 | if (initial_slow_test) { |
1355 | slow_region->init_req(need_gc_path, needgc_ctrl); |
1356 | // This completes all paths into the slow merge point |
1357 | transform_later(slow_region); |
1358 | } else { // No initial slow path needed! |
1359 | // Just fall from the need-GC path straight into the VM call. |
1360 | slow_region = needgc_ctrl; |
1361 | } |
1362 | |
1363 | InitializeNode* init = alloc->initialization(); |
1364 | fast_oop_rawmem = initialize_object(alloc, |
1365 | fast_oop_ctrl, fast_oop_rawmem, fast_oop, |
1366 | klass_node, length, size_in_bytes); |
1367 | |
1368 | // If initialization is performed by an array copy, any required |
1369 | // MemBarStoreStore was already added. If the object does not |
1370 | // escape no need for a MemBarStoreStore. If the object does not |
1371 | // escape in its initializer and memory barrier (MemBarStoreStore or |
1372 | // stronger) is already added at exit of initializer, also no need |
1373 | // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore |
1374 | // so that stores that initialize this object can't be reordered |
1375 | // with a subsequent store that makes this object accessible by |
1376 | // other threads. |
1377 | // Other threads include java threads and JVM internal threads |
1378 | // (for example concurrent GC threads). Current concurrent GC |
1379 | // implementation: CMS and G1 will not scan newly created object, |
1380 | // so it's safe to skip storestore barrier when allocation does |
1381 | // not escape. |
1382 | if (!alloc->does_not_escape_thread() && |
1383 | !alloc->is_allocation_MemBar_redundant() && |
1384 | (init == NULL || !init->is_complete_with_arraycopy())) { |
1385 | if (init == NULL || init->req() < InitializeNode::RawStores) { |
1386 | // No InitializeNode or no stores captured by zeroing |
1387 | // elimination. Simply add the MemBarStoreStore after object |
1388 | // initialization. |
1389 | MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); |
1390 | transform_later(mb); |
1391 | |
1392 | mb->init_req(TypeFunc::Memory, fast_oop_rawmem); |
1393 | mb->init_req(TypeFunc::Control, fast_oop_ctrl); |
1394 | fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control); |
1395 | transform_later(fast_oop_ctrl); |
1396 | fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory); |
1397 | transform_later(fast_oop_rawmem); |
1398 | } else { |
1399 | // Add the MemBarStoreStore after the InitializeNode so that |
1400 | // all stores performing the initialization that were moved |
1401 | // before the InitializeNode happen before the storestore |
1402 | // barrier. |
1403 | |
1404 | Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control); |
1405 | Node* init_mem = init->proj_out_or_null(TypeFunc::Memory); |
1406 | |
1407 | MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); |
1408 | transform_later(mb); |
1409 | |
1410 | Node* ctrl = new ProjNode(init,TypeFunc::Control); |
1411 | transform_later(ctrl); |
1412 | Node* mem = new ProjNode(init,TypeFunc::Memory); |
1413 | transform_later(mem); |
1414 | |
1415 | // The MemBarStoreStore depends on control and memory coming |
1416 | // from the InitializeNode |
1417 | mb->init_req(TypeFunc::Memory, mem); |
1418 | mb->init_req(TypeFunc::Control, ctrl); |
1419 | |
1420 | ctrl = new ProjNode(mb,TypeFunc::Control); |
1421 | transform_later(ctrl); |
1422 | mem = new ProjNode(mb,TypeFunc::Memory); |
1423 | transform_later(mem); |
1424 | |
1425 | // All nodes that depended on the InitializeNode for control |
1426 | // and memory must now depend on the MemBarNode that itself |
1427 | // depends on the InitializeNode |
1428 | if (init_ctrl != NULL) { |
1429 | _igvn.replace_node(init_ctrl, ctrl); |
1430 | } |
1431 | if (init_mem != NULL) { |
1432 | _igvn.replace_node(init_mem, mem); |
1433 | } |
1434 | } |
1435 | } |
1436 | |
1437 | if (C->env()->dtrace_extended_probes()) { |
1438 | // Slow-path call |
1439 | int size = TypeFunc::Parms + 2; |
1440 | CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), |
1441 | CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), |
1442 | "dtrace_object_alloc" , |
1443 | TypeRawPtr::BOTTOM); |
1444 | |
1445 | // Get base of thread-local storage area |
1446 | Node* thread = new ThreadLocalNode(); |
1447 | transform_later(thread); |
1448 | |
1449 | call->init_req(TypeFunc::Parms+0, thread); |
1450 | call->init_req(TypeFunc::Parms+1, fast_oop); |
1451 | call->init_req(TypeFunc::Control, fast_oop_ctrl); |
1452 | call->init_req(TypeFunc::I_O , top()); // does no i/o |
1453 | call->init_req(TypeFunc::Memory , fast_oop_rawmem); |
1454 | call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); |
1455 | call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); |
1456 | transform_later(call); |
1457 | fast_oop_ctrl = new ProjNode(call,TypeFunc::Control); |
1458 | transform_later(fast_oop_ctrl); |
1459 | fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory); |
1460 | transform_later(fast_oop_rawmem); |
1461 | } |
1462 | |
1463 | // Plug in the successful fast-path into the result merge point |
1464 | result_region ->init_req(fast_result_path, fast_oop_ctrl); |
1465 | result_phi_rawoop->init_req(fast_result_path, fast_oop); |
1466 | result_phi_i_o ->init_req(fast_result_path, i_o); |
1467 | result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); |
1468 | } else { |
1469 | slow_region = ctrl; |
1470 | result_phi_i_o = i_o; // Rename it to use in the following code. |
1471 | } |
1472 | |
1473 | // Generate slow-path call |
1474 | CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address, |
1475 | OptoRuntime::stub_name(slow_call_address), |
1476 | alloc->jvms()->bci(), |
1477 | TypePtr::BOTTOM); |
1478 | call->init_req( TypeFunc::Control, slow_region ); |
1479 | call->init_req( TypeFunc::I_O , top() ) ; // does no i/o |
1480 | call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs |
1481 | call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); |
1482 | call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); |
1483 | |
1484 | call->init_req(TypeFunc::Parms+0, klass_node); |
1485 | if (length != NULL) { |
1486 | call->init_req(TypeFunc::Parms+1, length); |
1487 | } |
1488 | |
1489 | // Copy debug information and adjust JVMState information, then replace |
1490 | // allocate node with the call |
1491 | copy_call_debug_info((CallNode *) alloc, call); |
1492 | if (!always_slow) { |
1493 | call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. |
1494 | } else { |
1495 | // Hook i_o projection to avoid its elimination during allocation |
1496 | // replacement (when only a slow call is generated). |
1497 | call->set_req(TypeFunc::I_O, result_phi_i_o); |
1498 | } |
1499 | _igvn.replace_node(alloc, call); |
1500 | transform_later(call); |
1501 | |
1502 | // Identify the output projections from the allocate node and |
1503 | // adjust any references to them. |
1504 | // The control and io projections look like: |
1505 | // |
1506 | // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) |
1507 | // Allocate Catch |
1508 | // ^---Proj(io) <-------+ ^---CatchProj(io) |
1509 | // |
1510 | // We are interested in the CatchProj nodes. |
1511 | // |
1512 | extract_call_projections(call); |
1513 | |
1514 | // An allocate node has separate memory projections for the uses on |
1515 | // the control and i_o paths. Replace the control memory projection with |
1516 | // result_phi_rawmem (unless we are only generating a slow call when |
1517 | // both memory projections are combined) |
1518 | if (!always_slow && _memproj_fallthrough != NULL) { |
1519 | for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { |
1520 | Node *use = _memproj_fallthrough->fast_out(i); |
1521 | _igvn.rehash_node_delayed(use); |
1522 | imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); |
1523 | // back up iterator |
1524 | --i; |
1525 | } |
1526 | } |
1527 | // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete |
1528 | // _memproj_catchall so we end up with a call that has only 1 memory projection. |
1529 | if (_memproj_catchall != NULL ) { |
1530 | if (_memproj_fallthrough == NULL) { |
1531 | _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory); |
1532 | transform_later(_memproj_fallthrough); |
1533 | } |
1534 | for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { |
1535 | Node *use = _memproj_catchall->fast_out(i); |
1536 | _igvn.rehash_node_delayed(use); |
1537 | imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); |
1538 | // back up iterator |
1539 | --i; |
1540 | } |
1541 | assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted" ); |
1542 | _igvn.remove_dead_node(_memproj_catchall); |
1543 | } |
1544 | |
1545 | // An allocate node has separate i_o projections for the uses on the control |
1546 | // and i_o paths. Always replace the control i_o projection with result i_o |
1547 | // otherwise incoming i_o become dead when only a slow call is generated |
1548 | // (it is different from memory projections where both projections are |
1549 | // combined in such case). |
1550 | if (_ioproj_fallthrough != NULL) { |
1551 | for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { |
1552 | Node *use = _ioproj_fallthrough->fast_out(i); |
1553 | _igvn.rehash_node_delayed(use); |
1554 | imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); |
1555 | // back up iterator |
1556 | --i; |
1557 | } |
1558 | } |
1559 | // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete |
1560 | // _ioproj_catchall so we end up with a call that has only 1 i_o projection. |
1561 | if (_ioproj_catchall != NULL ) { |
1562 | if (_ioproj_fallthrough == NULL) { |
1563 | _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O); |
1564 | transform_later(_ioproj_fallthrough); |
1565 | } |
1566 | for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { |
1567 | Node *use = _ioproj_catchall->fast_out(i); |
1568 | _igvn.rehash_node_delayed(use); |
1569 | imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); |
1570 | // back up iterator |
1571 | --i; |
1572 | } |
1573 | assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted" ); |
1574 | _igvn.remove_dead_node(_ioproj_catchall); |
1575 | } |
1576 | |
1577 | // if we generated only a slow call, we are done |
1578 | if (always_slow) { |
1579 | // Now we can unhook i_o. |
1580 | if (result_phi_i_o->outcnt() > 1) { |
1581 | call->set_req(TypeFunc::I_O, top()); |
1582 | } else { |
1583 | assert(result_phi_i_o->unique_ctrl_out() == call, "" ); |
1584 | // Case of new array with negative size known during compilation. |
1585 | // AllocateArrayNode::Ideal() optimization disconnect unreachable |
1586 | // following code since call to runtime will throw exception. |
1587 | // As result there will be no users of i_o after the call. |
1588 | // Leave i_o attached to this call to avoid problems in preceding graph. |
1589 | } |
1590 | return; |
1591 | } |
1592 | |
1593 | |
1594 | if (_fallthroughcatchproj != NULL) { |
1595 | ctrl = _fallthroughcatchproj->clone(); |
1596 | transform_later(ctrl); |
1597 | _igvn.replace_node(_fallthroughcatchproj, result_region); |
1598 | } else { |
1599 | ctrl = top(); |
1600 | } |
1601 | Node *slow_result; |
1602 | if (_resproj == NULL) { |
1603 | // no uses of the allocation result |
1604 | slow_result = top(); |
1605 | } else { |
1606 | slow_result = _resproj->clone(); |
1607 | transform_later(slow_result); |
1608 | _igvn.replace_node(_resproj, result_phi_rawoop); |
1609 | } |
1610 | |
1611 | // Plug slow-path into result merge point |
1612 | result_region ->init_req( slow_result_path, ctrl ); |
1613 | result_phi_rawoop->init_req( slow_result_path, slow_result); |
1614 | result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); |
1615 | transform_later(result_region); |
1616 | transform_later(result_phi_rawoop); |
1617 | transform_later(result_phi_rawmem); |
1618 | transform_later(result_phi_i_o); |
1619 | // This completes all paths into the result merge point |
1620 | } |
1621 | |
1622 | |
1623 | // Helper for PhaseMacroExpand::expand_allocate_common. |
1624 | // Initializes the newly-allocated storage. |
1625 | Node* |
1626 | PhaseMacroExpand::initialize_object(AllocateNode* alloc, |
1627 | Node* control, Node* rawmem, Node* object, |
1628 | Node* klass_node, Node* length, |
1629 | Node* size_in_bytes) { |
1630 | InitializeNode* init = alloc->initialization(); |
1631 | // Store the klass & mark bits |
1632 | Node* mark_node = NULL; |
1633 | // For now only enable fast locking for non-array types |
1634 | if (UseBiasedLocking && (length == NULL)) { |
1635 | mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); |
1636 | } else { |
1637 | mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); |
1638 | } |
1639 | rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); |
1640 | |
1641 | rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); |
1642 | int = alloc->minimum_header_size(); // conservatively small |
1643 | |
1644 | // Array length |
1645 | if (length != NULL) { // Arrays need length field |
1646 | rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); |
1647 | // conservatively small header size: |
1648 | header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
1649 | ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); |
1650 | if (k->is_array_klass()) // we know the exact header size in most cases: |
1651 | header_size = Klass::layout_helper_header_size(k->layout_helper()); |
1652 | } |
1653 | |
1654 | // Clear the object body, if necessary. |
1655 | if (init == NULL) { |
1656 | // The init has somehow disappeared; be cautious and clear everything. |
1657 | // |
1658 | // This can happen if a node is allocated but an uncommon trap occurs |
1659 | // immediately. In this case, the Initialize gets associated with the |
1660 | // trap, and may be placed in a different (outer) loop, if the Allocate |
1661 | // is in a loop. If (this is rare) the inner loop gets unrolled, then |
1662 | // there can be two Allocates to one Initialize. The answer in all these |
1663 | // edge cases is safety first. It is always safe to clear immediately |
1664 | // within an Allocate, and then (maybe or maybe not) clear some more later. |
1665 | if (!(UseTLAB && ZeroTLAB)) { |
1666 | rawmem = ClearArrayNode::clear_memory(control, rawmem, object, |
1667 | header_size, size_in_bytes, |
1668 | &_igvn); |
1669 | } |
1670 | } else { |
1671 | if (!init->is_complete()) { |
1672 | // Try to win by zeroing only what the init does not store. |
1673 | // We can also try to do some peephole optimizations, |
1674 | // such as combining some adjacent subword stores. |
1675 | rawmem = init->complete_stores(control, rawmem, object, |
1676 | header_size, size_in_bytes, &_igvn); |
1677 | } |
1678 | // We have no more use for this link, since the AllocateNode goes away: |
1679 | init->set_req(InitializeNode::RawAddress, top()); |
1680 | // (If we keep the link, it just confuses the register allocator, |
1681 | // who thinks he sees a real use of the address by the membar.) |
1682 | } |
1683 | |
1684 | return rawmem; |
1685 | } |
1686 | |
1687 | // Generate prefetch instructions for next allocations. |
1688 | Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, |
1689 | Node*& contended_phi_rawmem, |
1690 | Node* old_eden_top, Node* new_eden_top, |
1691 | intx lines) { |
1692 | enum { fall_in_path = 1, pf_path = 2 }; |
1693 | if( UseTLAB && AllocatePrefetchStyle == 2 ) { |
1694 | // Generate prefetch allocation with watermark check. |
1695 | // As an allocation hits the watermark, we will prefetch starting |
1696 | // at a "distance" away from watermark. |
1697 | |
1698 | Node *pf_region = new RegionNode(3); |
1699 | Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY, |
1700 | TypeRawPtr::BOTTOM ); |
1701 | // I/O is used for Prefetch |
1702 | Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO ); |
1703 | |
1704 | Node *thread = new ThreadLocalNode(); |
1705 | transform_later(thread); |
1706 | |
1707 | Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread, |
1708 | _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); |
1709 | transform_later(eden_pf_adr); |
1710 | |
1711 | Node *old_pf_wm = new LoadPNode(needgc_false, |
1712 | contended_phi_rawmem, eden_pf_adr, |
1713 | TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, |
1714 | MemNode::unordered); |
1715 | transform_later(old_pf_wm); |
1716 | |
1717 | // check against new_eden_top |
1718 | Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm ); |
1719 | transform_later(need_pf_cmp); |
1720 | Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge ); |
1721 | transform_later(need_pf_bol); |
1722 | IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol, |
1723 | PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); |
1724 | transform_later(need_pf_iff); |
1725 | |
1726 | // true node, add prefetchdistance |
1727 | Node *need_pf_true = new IfTrueNode( need_pf_iff ); |
1728 | transform_later(need_pf_true); |
1729 | |
1730 | Node *need_pf_false = new IfFalseNode( need_pf_iff ); |
1731 | transform_later(need_pf_false); |
1732 | |
1733 | Node *new_pf_wmt = new AddPNode( top(), old_pf_wm, |
1734 | _igvn.MakeConX(AllocatePrefetchDistance) ); |
1735 | transform_later(new_pf_wmt ); |
1736 | new_pf_wmt->set_req(0, need_pf_true); |
1737 | |
1738 | Node *store_new_wmt = new StorePNode(need_pf_true, |
1739 | contended_phi_rawmem, eden_pf_adr, |
1740 | TypeRawPtr::BOTTOM, new_pf_wmt, |
1741 | MemNode::unordered); |
1742 | transform_later(store_new_wmt); |
1743 | |
1744 | // adding prefetches |
1745 | pf_phi_abio->init_req( fall_in_path, i_o ); |
1746 | |
1747 | Node *prefetch_adr; |
1748 | Node *prefetch; |
1749 | uint step_size = AllocatePrefetchStepSize; |
1750 | uint distance = 0; |
1751 | |
1752 | for ( intx i = 0; i < lines; i++ ) { |
1753 | prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt, |
1754 | _igvn.MakeConX(distance) ); |
1755 | transform_later(prefetch_adr); |
1756 | prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); |
1757 | transform_later(prefetch); |
1758 | distance += step_size; |
1759 | i_o = prefetch; |
1760 | } |
1761 | pf_phi_abio->set_req( pf_path, i_o ); |
1762 | |
1763 | pf_region->init_req( fall_in_path, need_pf_false ); |
1764 | pf_region->init_req( pf_path, need_pf_true ); |
1765 | |
1766 | pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); |
1767 | pf_phi_rawmem->init_req( pf_path, store_new_wmt ); |
1768 | |
1769 | transform_later(pf_region); |
1770 | transform_later(pf_phi_rawmem); |
1771 | transform_later(pf_phi_abio); |
1772 | |
1773 | needgc_false = pf_region; |
1774 | contended_phi_rawmem = pf_phi_rawmem; |
1775 | i_o = pf_phi_abio; |
1776 | } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { |
1777 | // Insert a prefetch instruction for each allocation. |
1778 | // This code is used to generate 1 prefetch instruction per cache line. |
1779 | |
1780 | // Generate several prefetch instructions. |
1781 | uint step_size = AllocatePrefetchStepSize; |
1782 | uint distance = AllocatePrefetchDistance; |
1783 | |
1784 | // Next cache address. |
1785 | Node *cache_adr = new AddPNode(old_eden_top, old_eden_top, |
1786 | _igvn.MakeConX(step_size + distance)); |
1787 | transform_later(cache_adr); |
1788 | cache_adr = new CastP2XNode(needgc_false, cache_adr); |
1789 | transform_later(cache_adr); |
1790 | // Address is aligned to execute prefetch to the beginning of cache line size |
1791 | // (it is important when BIS instruction is used on SPARC as prefetch). |
1792 | Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); |
1793 | cache_adr = new AndXNode(cache_adr, mask); |
1794 | transform_later(cache_adr); |
1795 | cache_adr = new CastX2PNode(cache_adr); |
1796 | transform_later(cache_adr); |
1797 | |
1798 | // Prefetch |
1799 | Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); |
1800 | prefetch->set_req(0, needgc_false); |
1801 | transform_later(prefetch); |
1802 | contended_phi_rawmem = prefetch; |
1803 | Node *prefetch_adr; |
1804 | distance = step_size; |
1805 | for ( intx i = 1; i < lines; i++ ) { |
1806 | prefetch_adr = new AddPNode( cache_adr, cache_adr, |
1807 | _igvn.MakeConX(distance) ); |
1808 | transform_later(prefetch_adr); |
1809 | prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); |
1810 | transform_later(prefetch); |
1811 | distance += step_size; |
1812 | contended_phi_rawmem = prefetch; |
1813 | } |
1814 | } else if( AllocatePrefetchStyle > 0 ) { |
1815 | // Insert a prefetch for each allocation only on the fast-path |
1816 | Node *prefetch_adr; |
1817 | Node *prefetch; |
1818 | // Generate several prefetch instructions. |
1819 | uint step_size = AllocatePrefetchStepSize; |
1820 | uint distance = AllocatePrefetchDistance; |
1821 | for ( intx i = 0; i < lines; i++ ) { |
1822 | prefetch_adr = new AddPNode( old_eden_top, new_eden_top, |
1823 | _igvn.MakeConX(distance) ); |
1824 | transform_later(prefetch_adr); |
1825 | prefetch = new PrefetchAllocationNode( i_o, prefetch_adr ); |
1826 | // Do not let it float too high, since if eden_top == eden_end, |
1827 | // both might be null. |
1828 | if( i == 0 ) { // Set control for first prefetch, next follows it |
1829 | prefetch->init_req(0, needgc_false); |
1830 | } |
1831 | transform_later(prefetch); |
1832 | distance += step_size; |
1833 | i_o = prefetch; |
1834 | } |
1835 | } |
1836 | return i_o; |
1837 | } |
1838 | |
1839 | |
1840 | void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { |
1841 | expand_allocate_common(alloc, NULL, |
1842 | OptoRuntime::new_instance_Type(), |
1843 | OptoRuntime::new_instance_Java()); |
1844 | } |
1845 | |
1846 | void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { |
1847 | Node* length = alloc->in(AllocateNode::ALength); |
1848 | InitializeNode* init = alloc->initialization(); |
1849 | Node* klass_node = alloc->in(AllocateNode::KlassNode); |
1850 | ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); |
1851 | address slow_call_address; // Address of slow call |
1852 | if (init != NULL && init->is_complete_with_arraycopy() && |
1853 | k->is_type_array_klass()) { |
1854 | // Don't zero type array during slow allocation in VM since |
1855 | // it will be initialized later by arraycopy in compiled code. |
1856 | slow_call_address = OptoRuntime::new_array_nozero_Java(); |
1857 | } else { |
1858 | slow_call_address = OptoRuntime::new_array_Java(); |
1859 | } |
1860 | expand_allocate_common(alloc, length, |
1861 | OptoRuntime::new_array_Type(), |
1862 | slow_call_address); |
1863 | } |
1864 | |
1865 | //-------------------mark_eliminated_box---------------------------------- |
1866 | // |
1867 | // During EA obj may point to several objects but after few ideal graph |
1868 | // transformations (CCP) it may point to only one non escaping object |
1869 | // (but still using phi), corresponding locks and unlocks will be marked |
1870 | // for elimination. Later obj could be replaced with a new node (new phi) |
1871 | // and which does not have escape information. And later after some graph |
1872 | // reshape other locks and unlocks (which were not marked for elimination |
1873 | // before) are connected to this new obj (phi) but they still will not be |
1874 | // marked for elimination since new obj has no escape information. |
1875 | // Mark all associated (same box and obj) lock and unlock nodes for |
1876 | // elimination if some of them marked already. |
1877 | void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { |
1878 | if (oldbox->as_BoxLock()->is_eliminated()) |
1879 | return; // This BoxLock node was processed already. |
1880 | |
1881 | // New implementation (EliminateNestedLocks) has separate BoxLock |
1882 | // node for each locked region so mark all associated locks/unlocks as |
1883 | // eliminated even if different objects are referenced in one locked region |
1884 | // (for example, OSR compilation of nested loop inside locked scope). |
1885 | if (EliminateNestedLocks || |
1886 | oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { |
1887 | // Box is used only in one lock region. Mark this box as eliminated. |
1888 | _igvn.hash_delete(oldbox); |
1889 | oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value |
1890 | _igvn.hash_insert(oldbox); |
1891 | |
1892 | for (uint i = 0; i < oldbox->outcnt(); i++) { |
1893 | Node* u = oldbox->raw_out(i); |
1894 | if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { |
1895 | AbstractLockNode* alock = u->as_AbstractLock(); |
1896 | // Check lock's box since box could be referenced by Lock's debug info. |
1897 | if (alock->box_node() == oldbox) { |
1898 | // Mark eliminated all related locks and unlocks. |
1899 | #ifdef ASSERT |
1900 | alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4" ); |
1901 | #endif |
1902 | alock->set_non_esc_obj(); |
1903 | } |
1904 | } |
1905 | } |
1906 | return; |
1907 | } |
1908 | |
1909 | // Create new "eliminated" BoxLock node and use it in monitor debug info |
1910 | // instead of oldbox for the same object. |
1911 | BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); |
1912 | |
1913 | // Note: BoxLock node is marked eliminated only here and it is used |
1914 | // to indicate that all associated lock and unlock nodes are marked |
1915 | // for elimination. |
1916 | newbox->set_eliminated(); |
1917 | transform_later(newbox); |
1918 | |
1919 | // Replace old box node with new box for all users of the same object. |
1920 | for (uint i = 0; i < oldbox->outcnt();) { |
1921 | bool next_edge = true; |
1922 | |
1923 | Node* u = oldbox->raw_out(i); |
1924 | if (u->is_AbstractLock()) { |
1925 | AbstractLockNode* alock = u->as_AbstractLock(); |
1926 | if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { |
1927 | // Replace Box and mark eliminated all related locks and unlocks. |
1928 | #ifdef ASSERT |
1929 | alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5" ); |
1930 | #endif |
1931 | alock->set_non_esc_obj(); |
1932 | _igvn.rehash_node_delayed(alock); |
1933 | alock->set_box_node(newbox); |
1934 | next_edge = false; |
1935 | } |
1936 | } |
1937 | if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { |
1938 | FastLockNode* flock = u->as_FastLock(); |
1939 | assert(flock->box_node() == oldbox, "sanity" ); |
1940 | _igvn.rehash_node_delayed(flock); |
1941 | flock->set_box_node(newbox); |
1942 | next_edge = false; |
1943 | } |
1944 | |
1945 | // Replace old box in monitor debug info. |
1946 | if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { |
1947 | SafePointNode* sfn = u->as_SafePoint(); |
1948 | JVMState* youngest_jvms = sfn->jvms(); |
1949 | int max_depth = youngest_jvms->depth(); |
1950 | for (int depth = 1; depth <= max_depth; depth++) { |
1951 | JVMState* jvms = youngest_jvms->of_depth(depth); |
1952 | int num_mon = jvms->nof_monitors(); |
1953 | // Loop over monitors |
1954 | for (int idx = 0; idx < num_mon; idx++) { |
1955 | Node* obj_node = sfn->monitor_obj(jvms, idx); |
1956 | Node* box_node = sfn->monitor_box(jvms, idx); |
1957 | if (box_node == oldbox && obj_node->eqv_uncast(obj)) { |
1958 | int j = jvms->monitor_box_offset(idx); |
1959 | _igvn.replace_input_of(u, j, newbox); |
1960 | next_edge = false; |
1961 | } |
1962 | } |
1963 | } |
1964 | } |
1965 | if (next_edge) i++; |
1966 | } |
1967 | } |
1968 | |
1969 | //-----------------------mark_eliminated_locking_nodes----------------------- |
1970 | void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { |
1971 | if (EliminateNestedLocks) { |
1972 | if (alock->is_nested()) { |
1973 | assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity" ); |
1974 | return; |
1975 | } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened |
1976 | // Only Lock node has JVMState needed here. |
1977 | // Not that preceding claim is documented anywhere else. |
1978 | if (alock->jvms() != NULL) { |
1979 | if (alock->as_Lock()->is_nested_lock_region()) { |
1980 | // Mark eliminated related nested locks and unlocks. |
1981 | Node* obj = alock->obj_node(); |
1982 | BoxLockNode* box_node = alock->box_node()->as_BoxLock(); |
1983 | assert(!box_node->is_eliminated(), "should not be marked yet" ); |
1984 | // Note: BoxLock node is marked eliminated only here |
1985 | // and it is used to indicate that all associated lock |
1986 | // and unlock nodes are marked for elimination. |
1987 | box_node->set_eliminated(); // Box's hash is always NO_HASH here |
1988 | for (uint i = 0; i < box_node->outcnt(); i++) { |
1989 | Node* u = box_node->raw_out(i); |
1990 | if (u->is_AbstractLock()) { |
1991 | alock = u->as_AbstractLock(); |
1992 | if (alock->box_node() == box_node) { |
1993 | // Verify that this Box is referenced only by related locks. |
1994 | assert(alock->obj_node()->eqv_uncast(obj), "" ); |
1995 | // Mark all related locks and unlocks. |
1996 | #ifdef ASSERT |
1997 | alock->log_lock_optimization(C, "eliminate_lock_set_nested" ); |
1998 | #endif |
1999 | alock->set_nested(); |
2000 | } |
2001 | } |
2002 | } |
2003 | } else { |
2004 | #ifdef ASSERT |
2005 | alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region" ); |
2006 | if (C->log() != NULL) |
2007 | alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output |
2008 | #endif |
2009 | } |
2010 | } |
2011 | return; |
2012 | } |
2013 | // Process locks for non escaping object |
2014 | assert(alock->is_non_esc_obj(), "" ); |
2015 | } // EliminateNestedLocks |
2016 | |
2017 | if (alock->is_non_esc_obj()) { // Lock is used for non escaping object |
2018 | // Look for all locks of this object and mark them and |
2019 | // corresponding BoxLock nodes as eliminated. |
2020 | Node* obj = alock->obj_node(); |
2021 | for (uint j = 0; j < obj->outcnt(); j++) { |
2022 | Node* o = obj->raw_out(j); |
2023 | if (o->is_AbstractLock() && |
2024 | o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { |
2025 | alock = o->as_AbstractLock(); |
2026 | Node* box = alock->box_node(); |
2027 | // Replace old box node with new eliminated box for all users |
2028 | // of the same object and mark related locks as eliminated. |
2029 | mark_eliminated_box(box, obj); |
2030 | } |
2031 | } |
2032 | } |
2033 | } |
2034 | |
2035 | // we have determined that this lock/unlock can be eliminated, we simply |
2036 | // eliminate the node without expanding it. |
2037 | // |
2038 | // Note: The membar's associated with the lock/unlock are currently not |
2039 | // eliminated. This should be investigated as a future enhancement. |
2040 | // |
2041 | bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { |
2042 | |
2043 | if (!alock->is_eliminated()) { |
2044 | return false; |
2045 | } |
2046 | #ifdef ASSERT |
2047 | if (!alock->is_coarsened()) { |
2048 | // Check that new "eliminated" BoxLock node is created. |
2049 | BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); |
2050 | assert(oldbox->is_eliminated(), "should be done already" ); |
2051 | } |
2052 | #endif |
2053 | |
2054 | alock->log_lock_optimization(C, "eliminate_lock" ); |
2055 | |
2056 | #ifndef PRODUCT |
2057 | if (PrintEliminateLocks) { |
2058 | if (alock->is_Lock()) { |
2059 | tty->print_cr("++++ Eliminated: %d Lock" , alock->_idx); |
2060 | } else { |
2061 | tty->print_cr("++++ Eliminated: %d Unlock" , alock->_idx); |
2062 | } |
2063 | } |
2064 | #endif |
2065 | |
2066 | Node* mem = alock->in(TypeFunc::Memory); |
2067 | Node* ctrl = alock->in(TypeFunc::Control); |
2068 | guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL" ); |
2069 | |
2070 | extract_call_projections(alock); |
2071 | // There are 2 projections from the lock. The lock node will |
2072 | // be deleted when its last use is subsumed below. |
2073 | assert(alock->outcnt() == 2 && |
2074 | _fallthroughproj != NULL && |
2075 | _memproj_fallthrough != NULL, |
2076 | "Unexpected projections from Lock/Unlock" ); |
2077 | |
2078 | Node* fallthroughproj = _fallthroughproj; |
2079 | Node* memproj_fallthrough = _memproj_fallthrough; |
2080 | |
2081 | // The memory projection from a lock/unlock is RawMem |
2082 | // The input to a Lock is merged memory, so extract its RawMem input |
2083 | // (unless the MergeMem has been optimized away.) |
2084 | if (alock->is_Lock()) { |
2085 | // Seach for MemBarAcquireLock node and delete it also. |
2086 | MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); |
2087 | assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "" ); |
2088 | Node* ctrlproj = membar->proj_out(TypeFunc::Control); |
2089 | Node* memproj = membar->proj_out(TypeFunc::Memory); |
2090 | _igvn.replace_node(ctrlproj, fallthroughproj); |
2091 | _igvn.replace_node(memproj, memproj_fallthrough); |
2092 | |
2093 | // Delete FastLock node also if this Lock node is unique user |
2094 | // (a loop peeling may clone a Lock node). |
2095 | Node* flock = alock->as_Lock()->fastlock_node(); |
2096 | if (flock->outcnt() == 1) { |
2097 | assert(flock->unique_out() == alock, "sanity" ); |
2098 | _igvn.replace_node(flock, top()); |
2099 | } |
2100 | } |
2101 | |
2102 | // Seach for MemBarReleaseLock node and delete it also. |
2103 | if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) { |
2104 | MemBarNode* membar = ctrl->in(0)->as_MemBar(); |
2105 | assert(membar->Opcode() == Op_MemBarReleaseLock && |
2106 | mem->is_Proj() && membar == mem->in(0), "" ); |
2107 | _igvn.replace_node(fallthroughproj, ctrl); |
2108 | _igvn.replace_node(memproj_fallthrough, mem); |
2109 | fallthroughproj = ctrl; |
2110 | memproj_fallthrough = mem; |
2111 | ctrl = membar->in(TypeFunc::Control); |
2112 | mem = membar->in(TypeFunc::Memory); |
2113 | } |
2114 | |
2115 | _igvn.replace_node(fallthroughproj, ctrl); |
2116 | _igvn.replace_node(memproj_fallthrough, mem); |
2117 | return true; |
2118 | } |
2119 | |
2120 | |
2121 | //------------------------------expand_lock_node---------------------- |
2122 | void PhaseMacroExpand::expand_lock_node(LockNode *lock) { |
2123 | |
2124 | Node* ctrl = lock->in(TypeFunc::Control); |
2125 | Node* mem = lock->in(TypeFunc::Memory); |
2126 | Node* obj = lock->obj_node(); |
2127 | Node* box = lock->box_node(); |
2128 | Node* flock = lock->fastlock_node(); |
2129 | |
2130 | assert(!box->as_BoxLock()->is_eliminated(), "sanity" ); |
2131 | |
2132 | // Make the merge point |
2133 | Node *region; |
2134 | Node *mem_phi; |
2135 | Node *slow_path; |
2136 | |
2137 | if (UseOptoBiasInlining) { |
2138 | /* |
2139 | * See the full description in MacroAssembler::biased_locking_enter(). |
2140 | * |
2141 | * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { |
2142 | * // The object is biased. |
2143 | * proto_node = klass->prototype_header; |
2144 | * o_node = thread | proto_node; |
2145 | * x_node = o_node ^ mark_word; |
2146 | * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? |
2147 | * // Done. |
2148 | * } else { |
2149 | * if( (x_node & biased_lock_mask) != 0 ) { |
2150 | * // The klass's prototype header is no longer biased. |
2151 | * cas(&mark_word, mark_word, proto_node) |
2152 | * goto cas_lock; |
2153 | * } else { |
2154 | * // The klass's prototype header is still biased. |
2155 | * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? |
2156 | * old = mark_word; |
2157 | * new = o_node; |
2158 | * } else { |
2159 | * // Different thread or anonymous biased. |
2160 | * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); |
2161 | * new = thread | old; |
2162 | * } |
2163 | * // Try to rebias. |
2164 | * if( cas(&mark_word, old, new) == 0 ) { |
2165 | * // Done. |
2166 | * } else { |
2167 | * goto slow_path; // Failed. |
2168 | * } |
2169 | * } |
2170 | * } |
2171 | * } else { |
2172 | * // The object is not biased. |
2173 | * cas_lock: |
2174 | * if( FastLock(obj) == 0 ) { |
2175 | * // Done. |
2176 | * } else { |
2177 | * slow_path: |
2178 | * OptoRuntime::complete_monitor_locking_Java(obj); |
2179 | * } |
2180 | * } |
2181 | */ |
2182 | |
2183 | region = new RegionNode(5); |
2184 | // create a Phi for the memory state |
2185 | mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
2186 | |
2187 | Node* fast_lock_region = new RegionNode(3); |
2188 | Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); |
2189 | |
2190 | // First, check mark word for the biased lock pattern. |
2191 | Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); |
2192 | |
2193 | // Get fast path - mark word has the biased lock pattern. |
2194 | ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, |
2195 | markOopDesc::biased_lock_mask_in_place, |
2196 | markOopDesc::biased_lock_pattern, true); |
2197 | // fast_lock_region->in(1) is set to slow path. |
2198 | fast_lock_mem_phi->init_req(1, mem); |
2199 | |
2200 | // Now check that the lock is biased to the current thread and has |
2201 | // the same epoch and bias as Klass::_prototype_header. |
2202 | |
2203 | // Special-case a fresh allocation to avoid building nodes: |
2204 | Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); |
2205 | if (klass_node == NULL) { |
2206 | Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); |
2207 | klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr())); |
2208 | #ifdef _LP64 |
2209 | if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { |
2210 | assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity" ); |
2211 | klass_node->in(1)->init_req(0, ctrl); |
2212 | } else |
2213 | #endif |
2214 | klass_node->init_req(0, ctrl); |
2215 | } |
2216 | Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); |
2217 | |
2218 | Node* thread = transform_later(new ThreadLocalNode()); |
2219 | Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread)); |
2220 | Node* o_node = transform_later(new OrXNode(cast_thread, proto_node)); |
2221 | Node* x_node = transform_later(new XorXNode(o_node, mark_node)); |
2222 | |
2223 | // Get slow path - mark word does NOT match the value. |
2224 | Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, |
2225 | (~markOopDesc::age_mask_in_place), 0); |
2226 | // region->in(3) is set to fast path - the object is biased to the current thread. |
2227 | mem_phi->init_req(3, mem); |
2228 | |
2229 | |
2230 | // Mark word does NOT match the value (thread | Klass::_prototype_header). |
2231 | |
2232 | |
2233 | // First, check biased pattern. |
2234 | // Get fast path - _prototype_header has the same biased lock pattern. |
2235 | ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, |
2236 | markOopDesc::biased_lock_mask_in_place, 0, true); |
2237 | |
2238 | not_biased_ctrl = fast_lock_region->in(2); // Slow path |
2239 | // fast_lock_region->in(2) - the prototype header is no longer biased |
2240 | // and we have to revoke the bias on this object. |
2241 | // We are going to try to reset the mark of this object to the prototype |
2242 | // value and fall through to the CAS-based locking scheme. |
2243 | Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); |
2244 | Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr, |
2245 | proto_node, mark_node); |
2246 | transform_later(cas); |
2247 | Node* proj = transform_later(new SCMemProjNode(cas)); |
2248 | fast_lock_mem_phi->init_req(2, proj); |
2249 | |
2250 | |
2251 | // Second, check epoch bits. |
2252 | Node* rebiased_region = new RegionNode(3); |
2253 | Node* old_phi = new PhiNode( rebiased_region, TypeX_X); |
2254 | Node* new_phi = new PhiNode( rebiased_region, TypeX_X); |
2255 | |
2256 | // Get slow path - mark word does NOT match epoch bits. |
2257 | Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, |
2258 | markOopDesc::epoch_mask_in_place, 0); |
2259 | // The epoch of the current bias is not valid, attempt to rebias the object |
2260 | // toward the current thread. |
2261 | rebiased_region->init_req(2, epoch_ctrl); |
2262 | old_phi->init_req(2, mark_node); |
2263 | new_phi->init_req(2, o_node); |
2264 | |
2265 | // rebiased_region->in(1) is set to fast path. |
2266 | // The epoch of the current bias is still valid but we know |
2267 | // nothing about the owner; it might be set or it might be clear. |
2268 | Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | |
2269 | markOopDesc::age_mask_in_place | |
2270 | markOopDesc::epoch_mask_in_place); |
2271 | Node* old = transform_later(new AndXNode(mark_node, cmask)); |
2272 | cast_thread = transform_later(new CastP2XNode(ctrl, thread)); |
2273 | Node* new_mark = transform_later(new OrXNode(cast_thread, old)); |
2274 | old_phi->init_req(1, old); |
2275 | new_phi->init_req(1, new_mark); |
2276 | |
2277 | transform_later(rebiased_region); |
2278 | transform_later(old_phi); |
2279 | transform_later(new_phi); |
2280 | |
2281 | // Try to acquire the bias of the object using an atomic operation. |
2282 | // If this fails we will go in to the runtime to revoke the object's bias. |
2283 | cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi); |
2284 | transform_later(cas); |
2285 | proj = transform_later(new SCMemProjNode(cas)); |
2286 | |
2287 | // Get slow path - Failed to CAS. |
2288 | not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); |
2289 | mem_phi->init_req(4, proj); |
2290 | // region->in(4) is set to fast path - the object is rebiased to the current thread. |
2291 | |
2292 | // Failed to CAS. |
2293 | slow_path = new RegionNode(3); |
2294 | Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); |
2295 | |
2296 | slow_path->init_req(1, not_biased_ctrl); // Capture slow-control |
2297 | slow_mem->init_req(1, proj); |
2298 | |
2299 | // Call CAS-based locking scheme (FastLock node). |
2300 | |
2301 | transform_later(fast_lock_region); |
2302 | transform_later(fast_lock_mem_phi); |
2303 | |
2304 | // Get slow path - FastLock failed to lock the object. |
2305 | ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); |
2306 | mem_phi->init_req(2, fast_lock_mem_phi); |
2307 | // region->in(2) is set to fast path - the object is locked to the current thread. |
2308 | |
2309 | slow_path->init_req(2, ctrl); // Capture slow-control |
2310 | slow_mem->init_req(2, fast_lock_mem_phi); |
2311 | |
2312 | transform_later(slow_path); |
2313 | transform_later(slow_mem); |
2314 | // Reset lock's memory edge. |
2315 | lock->set_req(TypeFunc::Memory, slow_mem); |
2316 | |
2317 | } else { |
2318 | region = new RegionNode(3); |
2319 | // create a Phi for the memory state |
2320 | mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
2321 | |
2322 | // Optimize test; set region slot 2 |
2323 | slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); |
2324 | mem_phi->init_req(2, mem); |
2325 | } |
2326 | |
2327 | // Make slow path call |
2328 | CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), |
2329 | OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, |
2330 | obj, box, NULL); |
2331 | |
2332 | extract_call_projections(call); |
2333 | |
2334 | // Slow path can only throw asynchronous exceptions, which are always |
2335 | // de-opted. So the compiler thinks the slow-call can never throw an |
2336 | // exception. If it DOES throw an exception we would need the debug |
2337 | // info removed first (since if it throws there is no monitor). |
2338 | assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && |
2339 | _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock" ); |
2340 | |
2341 | // Capture slow path |
2342 | // disconnect fall-through projection from call and create a new one |
2343 | // hook up users of fall-through projection to region |
2344 | Node *slow_ctrl = _fallthroughproj->clone(); |
2345 | transform_later(slow_ctrl); |
2346 | _igvn.hash_delete(_fallthroughproj); |
2347 | _fallthroughproj->disconnect_inputs(NULL, C); |
2348 | region->init_req(1, slow_ctrl); |
2349 | // region inputs are now complete |
2350 | transform_later(region); |
2351 | _igvn.replace_node(_fallthroughproj, region); |
2352 | |
2353 | Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory)); |
2354 | mem_phi->init_req(1, memproj ); |
2355 | transform_later(mem_phi); |
2356 | _igvn.replace_node(_memproj_fallthrough, mem_phi); |
2357 | } |
2358 | |
2359 | //------------------------------expand_unlock_node---------------------- |
2360 | void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { |
2361 | |
2362 | Node* ctrl = unlock->in(TypeFunc::Control); |
2363 | Node* mem = unlock->in(TypeFunc::Memory); |
2364 | Node* obj = unlock->obj_node(); |
2365 | Node* box = unlock->box_node(); |
2366 | |
2367 | assert(!box->as_BoxLock()->is_eliminated(), "sanity" ); |
2368 | |
2369 | // No need for a null check on unlock |
2370 | |
2371 | // Make the merge point |
2372 | Node *region; |
2373 | Node *mem_phi; |
2374 | |
2375 | if (UseOptoBiasInlining) { |
2376 | // Check for biased locking unlock case, which is a no-op. |
2377 | // See the full description in MacroAssembler::biased_locking_exit(). |
2378 | region = new RegionNode(4); |
2379 | // create a Phi for the memory state |
2380 | mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
2381 | mem_phi->init_req(3, mem); |
2382 | |
2383 | Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); |
2384 | ctrl = opt_bits_test(ctrl, region, 3, mark_node, |
2385 | markOopDesc::biased_lock_mask_in_place, |
2386 | markOopDesc::biased_lock_pattern); |
2387 | } else { |
2388 | region = new RegionNode(3); |
2389 | // create a Phi for the memory state |
2390 | mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
2391 | } |
2392 | |
2393 | FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box ); |
2394 | funlock = transform_later( funlock )->as_FastUnlock(); |
2395 | // Optimize test; set region slot 2 |
2396 | Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); |
2397 | Node *thread = transform_later(new ThreadLocalNode()); |
2398 | |
2399 | CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), |
2400 | CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), |
2401 | "complete_monitor_unlocking_C" , slow_path, obj, box, thread); |
2402 | |
2403 | extract_call_projections(call); |
2404 | |
2405 | assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && |
2406 | _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock" ); |
2407 | |
2408 | // No exceptions for unlocking |
2409 | // Capture slow path |
2410 | // disconnect fall-through projection from call and create a new one |
2411 | // hook up users of fall-through projection to region |
2412 | Node *slow_ctrl = _fallthroughproj->clone(); |
2413 | transform_later(slow_ctrl); |
2414 | _igvn.hash_delete(_fallthroughproj); |
2415 | _fallthroughproj->disconnect_inputs(NULL, C); |
2416 | region->init_req(1, slow_ctrl); |
2417 | // region inputs are now complete |
2418 | transform_later(region); |
2419 | _igvn.replace_node(_fallthroughproj, region); |
2420 | |
2421 | Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) ); |
2422 | mem_phi->init_req(1, memproj ); |
2423 | mem_phi->init_req(2, mem); |
2424 | transform_later(mem_phi); |
2425 | _igvn.replace_node(_memproj_fallthrough, mem_phi); |
2426 | } |
2427 | |
2428 | //---------------------------eliminate_macro_nodes---------------------- |
2429 | // Eliminate scalar replaced allocations and associated locks. |
2430 | void PhaseMacroExpand::eliminate_macro_nodes() { |
2431 | if (C->macro_count() == 0) |
2432 | return; |
2433 | |
2434 | // First, attempt to eliminate locks |
2435 | int cnt = C->macro_count(); |
2436 | for (int i=0; i < cnt; i++) { |
2437 | Node *n = C->macro_node(i); |
2438 | if (n->is_AbstractLock()) { // Lock and Unlock nodes |
2439 | // Before elimination mark all associated (same box and obj) |
2440 | // lock and unlock nodes. |
2441 | mark_eliminated_locking_nodes(n->as_AbstractLock()); |
2442 | } |
2443 | } |
2444 | bool progress = true; |
2445 | while (progress) { |
2446 | progress = false; |
2447 | for (int i = C->macro_count(); i > 0; i--) { |
2448 | Node * n = C->macro_node(i-1); |
2449 | bool success = false; |
2450 | debug_only(int old_macro_count = C->macro_count();); |
2451 | if (n->is_AbstractLock()) { |
2452 | success = eliminate_locking_node(n->as_AbstractLock()); |
2453 | } |
2454 | assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count" ); |
2455 | progress = progress || success; |
2456 | } |
2457 | } |
2458 | // Next, attempt to eliminate allocations |
2459 | _has_locks = false; |
2460 | progress = true; |
2461 | while (progress) { |
2462 | progress = false; |
2463 | for (int i = C->macro_count(); i > 0; i--) { |
2464 | Node * n = C->macro_node(i-1); |
2465 | bool success = false; |
2466 | debug_only(int old_macro_count = C->macro_count();); |
2467 | switch (n->class_id()) { |
2468 | case Node::Class_Allocate: |
2469 | case Node::Class_AllocateArray: |
2470 | success = eliminate_allocate_node(n->as_Allocate()); |
2471 | break; |
2472 | case Node::Class_CallStaticJava: |
2473 | success = eliminate_boxing_node(n->as_CallStaticJava()); |
2474 | break; |
2475 | case Node::Class_Lock: |
2476 | case Node::Class_Unlock: |
2477 | assert(!n->as_AbstractLock()->is_eliminated(), "sanity" ); |
2478 | _has_locks = true; |
2479 | break; |
2480 | case Node::Class_ArrayCopy: |
2481 | break; |
2482 | case Node::Class_OuterStripMinedLoop: |
2483 | break; |
2484 | default: |
2485 | assert(n->Opcode() == Op_LoopLimit || |
2486 | n->Opcode() == Op_Opaque1 || |
2487 | n->Opcode() == Op_Opaque2 || |
2488 | n->Opcode() == Op_Opaque3 || |
2489 | BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n), |
2490 | "unknown node type in macro list" ); |
2491 | } |
2492 | assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count" ); |
2493 | progress = progress || success; |
2494 | } |
2495 | } |
2496 | } |
2497 | |
2498 | //------------------------------expand_macro_nodes---------------------- |
2499 | // Returns true if a failure occurred. |
2500 | bool PhaseMacroExpand::expand_macro_nodes() { |
2501 | // Last attempt to eliminate macro nodes. |
2502 | eliminate_macro_nodes(); |
2503 | |
2504 | // Make sure expansion will not cause node limit to be exceeded. |
2505 | // Worst case is a macro node gets expanded into about 200 nodes. |
2506 | // Allow 50% more for optimization. |
2507 | if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) ) |
2508 | return true; |
2509 | |
2510 | // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. |
2511 | bool progress = true; |
2512 | while (progress) { |
2513 | progress = false; |
2514 | for (int i = C->macro_count(); i > 0; i--) { |
2515 | Node * n = C->macro_node(i-1); |
2516 | bool success = false; |
2517 | debug_only(int old_macro_count = C->macro_count();); |
2518 | if (n->Opcode() == Op_LoopLimit) { |
2519 | // Remove it from macro list and put on IGVN worklist to optimize. |
2520 | C->remove_macro_node(n); |
2521 | _igvn._worklist.push(n); |
2522 | success = true; |
2523 | } else if (n->Opcode() == Op_CallStaticJava) { |
2524 | // Remove it from macro list and put on IGVN worklist to optimize. |
2525 | C->remove_macro_node(n); |
2526 | _igvn._worklist.push(n); |
2527 | success = true; |
2528 | } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { |
2529 | _igvn.replace_node(n, n->in(1)); |
2530 | success = true; |
2531 | #if INCLUDE_RTM_OPT |
2532 | } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) { |
2533 | assert(C->profile_rtm(), "should be used only in rtm deoptimization code" ); |
2534 | assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "" ); |
2535 | Node* cmp = n->unique_out(); |
2536 | #ifdef ASSERT |
2537 | // Validate graph. |
2538 | assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "" ); |
2539 | BoolNode* bol = cmp->unique_out()->as_Bool(); |
2540 | assert((bol->outcnt() == 1) && bol->unique_out()->is_If() && |
2541 | (bol->_test._test == BoolTest::ne), "" ); |
2542 | IfNode* ifn = bol->unique_out()->as_If(); |
2543 | assert((ifn->outcnt() == 2) && |
2544 | ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "" ); |
2545 | #endif |
2546 | Node* repl = n->in(1); |
2547 | if (!_has_locks) { |
2548 | // Remove RTM state check if there are no locks in the code. |
2549 | // Replace input to compare the same value. |
2550 | repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1); |
2551 | } |
2552 | _igvn.replace_node(n, repl); |
2553 | success = true; |
2554 | #endif |
2555 | } else if (n->Opcode() == Op_OuterStripMinedLoop) { |
2556 | n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn); |
2557 | C->remove_macro_node(n); |
2558 | success = true; |
2559 | } |
2560 | assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count" ); |
2561 | progress = progress || success; |
2562 | } |
2563 | } |
2564 | |
2565 | // expand arraycopy "macro" nodes first |
2566 | // For ReduceBulkZeroing, we must first process all arraycopy nodes |
2567 | // before the allocate nodes are expanded. |
2568 | int macro_idx = C->macro_count() - 1; |
2569 | while (macro_idx >= 0) { |
2570 | Node * n = C->macro_node(macro_idx); |
2571 | assert(n->is_macro(), "only macro nodes expected here" ); |
2572 | if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { |
2573 | // node is unreachable, so don't try to expand it |
2574 | C->remove_macro_node(n); |
2575 | } else if (n->is_ArrayCopy()){ |
2576 | int macro_count = C->macro_count(); |
2577 | expand_arraycopy_node(n->as_ArrayCopy()); |
2578 | assert(C->macro_count() < macro_count, "must have deleted a node from macro list" ); |
2579 | } |
2580 | if (C->failing()) return true; |
2581 | macro_idx --; |
2582 | } |
2583 | |
2584 | // expand "macro" nodes |
2585 | // nodes are removed from the macro list as they are processed |
2586 | while (C->macro_count() > 0) { |
2587 | int macro_count = C->macro_count(); |
2588 | Node * n = C->macro_node(macro_count-1); |
2589 | assert(n->is_macro(), "only macro nodes expected here" ); |
2590 | if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) { |
2591 | // node is unreachable, so don't try to expand it |
2592 | C->remove_macro_node(n); |
2593 | continue; |
2594 | } |
2595 | switch (n->class_id()) { |
2596 | case Node::Class_Allocate: |
2597 | expand_allocate(n->as_Allocate()); |
2598 | break; |
2599 | case Node::Class_AllocateArray: |
2600 | expand_allocate_array(n->as_AllocateArray()); |
2601 | break; |
2602 | case Node::Class_Lock: |
2603 | expand_lock_node(n->as_Lock()); |
2604 | break; |
2605 | case Node::Class_Unlock: |
2606 | expand_unlock_node(n->as_Unlock()); |
2607 | break; |
2608 | default: |
2609 | assert(false, "unknown node type in macro list" ); |
2610 | } |
2611 | assert(C->macro_count() < macro_count, "must have deleted a node from macro list" ); |
2612 | if (C->failing()) return true; |
2613 | } |
2614 | |
2615 | _igvn.set_delay_transform(false); |
2616 | _igvn.optimize(); |
2617 | if (C->failing()) return true; |
2618 | return false; |
2619 | } |
2620 | |